diff --git a/fr/1120ms/decoder.mlmodelc/analytics/coremldata.bin b/fr/1120ms/decoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..65b800e186dcb664ec3d46edb11fa43215dce1f7
--- /dev/null
+++ b/fr/1120ms/decoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d3becf0abd7a7efd30577f33f4144cfcb8c08155fdd62903eb1a45d7b02b87c
+size 243
diff --git a/fr/1120ms/decoder.mlmodelc/coremldata.bin b/fr/1120ms/decoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..a921778bfd0fef1f29bd148ddcb6c74c6116663e
--- /dev/null
+++ b/fr/1120ms/decoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d6b7630abf0c1b447913bf9d1ed5935d625bcb70e978c607820b2ff35ee8f07f
+size 433
diff --git a/fr/1120ms/decoder.mlmodelc/model.mil b/fr/1120ms/decoder.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..0de41956c1a2cac40f0629b26d6b60b23d5fe6b3
--- /dev/null
+++ b/fr/1120ms/decoder.mlmodelc/model.mil
@@ -0,0 +1,64 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            int32 y_axis_0 = const()[name = string("y_axis_0"), val = int32(0)];
+            int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)];
+            bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [849, 640]> module_prediction_embed_weight_to_fp16 = const()[name = string("module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_8")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)];
+            int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)];
+            string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")];
+            int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)];
+            int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)];
+            string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            string input_lstm_layer_0_direction_0 = const()[name = string("input_lstm_layer_0_direction_0"), val = string("forward")];
+            bool input_lstm_layer_0_output_sequence_0 = const()[name = string("input_lstm_layer_0_output_sequence_0"), val = bool(true)];
+            string input_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")];
+            string input_lstm_layer_0_cell_activation_0 = const()[name = string("input_lstm_layer_0_cell_activation_0"), val = string("tanh")];
+            string input_lstm_layer_0_activation_0 = const()[name = string("input_lstm_layer_0_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1086848)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4363712)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7640576)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = string("transpose_2")];
+            tensor<fp16, [1, 1, 640]> input_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_lstm_layer_0_cast_fp16_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_lstm_layer_0_cell_activation_0, direction = input_lstm_layer_0_direction_0, initial_c = input_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_lstm_c0_squeeze_cast_fp16")];
+            string input_direction_0 = const()[name = string("input_direction_0"), val = string("forward")];
+            bool input_output_sequence_0 = const()[name = string("input_output_sequence_0"), val = bool(true)];
+            string input_recurrent_activation_0 = const()[name = string("input_recurrent_activation_0"), val = string("sigmoid")];
+            string input_cell_activation_0 = const()[name = string("input_cell_activation_0"), val = string("tanh")];
+            string input_activation_0 = const()[name = string("input_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7645760)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10922624)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14199488)))];
+            tensor<fp16, [1, 1, 640]> input_cast_fp16_0, tensor<fp16, [1, 640]> input_cast_fp16_1, tensor<fp16, [1, 640]> input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_3_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_lstm_layer_0_cast_fp16_0)[name = string("input_cast_fp16")];
+            int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_cast_fp16_1, input_cast_fp16_1))[name = string("obj_3_cast_fp16")];
+            string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_lstm_layer_0_cast_fp16_2, input_cast_fp16_2))[name = string("obj_cast_fp16")];
+            string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
+            string transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = string("transpose_0_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp16, [1, 640, 1]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = string("transpose_1")];
+            tensor<fp32, [1, 640, 1]> decoder_out = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = string("cast_3")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")];
+        } -> (decoder_out, h_out, c_out);
+}
\ No newline at end of file
diff --git a/fr/1120ms/decoder.mlmodelc/weights/weight.bin b/fr/1120ms/decoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..82ea1696ddad052267574c52c3e3b0bccde479f5
--- /dev/null
+++ b/fr/1120ms/decoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ef2aeef85ed664eb60850330a014c5c299325ba1047d690654dc7ffa870e95a
+size 14204672
diff --git a/fr/1120ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/1120ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..c09a4248bce6ae30d0023f4ec349ae462e19aad7
--- /dev/null
+++ b/fr/1120ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c82c0fbdef3af45d82f9b1fe84c0c822154020a5b3cd0e5070d33e9ef0beeca6
+size 10359
diff --git a/fr/1120ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/1120ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..82ea1696ddad052267574c52c3e3b0bccde479f5
--- /dev/null
+++ b/fr/1120ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ef2aeef85ed664eb60850330a014c5c299325ba1047d690654dc7ffa870e95a
+size 14204672
diff --git a/fr/1120ms/decoder.mlpackage/Manifest.json b/fr/1120ms/decoder.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..80bff9abaf12e957f68dbde257eaf82079fed11b
--- /dev/null
+++ b/fr/1120ms/decoder.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "08617906-EB26-428B-BCF5-BC6013603E2C": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "89EF9907-E7AF-4633-B6D7-C8D472B636BC": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "89EF9907-E7AF-4633-B6D7-C8D472B636BC"
+}
diff --git a/fr/1120ms/decoder_joint.mlmodelc/analytics/coremldata.bin b/fr/1120ms/decoder_joint.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..22debc561a233f97bd7fde99dc9d1a31fb0231de
--- /dev/null
+++ b/fr/1120ms/decoder_joint.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f8367de6fb06c7b8053f5ecbbc8e4c7e5100f85f1e1c9ba6e9d9b3e31ecca3d
+size 243
diff --git a/fr/1120ms/decoder_joint.mlmodelc/coremldata.bin b/fr/1120ms/decoder_joint.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..54e005ff86e2274927ae7ef36a7e0ee4bb128c3b
--- /dev/null
+++ b/fr/1120ms/decoder_joint.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0566656fc89cd81234bb22159c3ccebfb8c0277e3b740d12a8fb53288c1efcd
+size 454
diff --git a/fr/1120ms/decoder_joint.mlmodelc/model.mil b/fr/1120ms/decoder_joint.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..1438ac2ef0ec1f9fc664a6752938e369f7de8e9d
--- /dev/null
+++ b/fr/1120ms/decoder_joint.mlmodelc/model.mil
@@ -0,0 +1,83 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [1, 1024, 1]> encoder, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            int32 y_axis_0 = const()[name = string("y_axis_0"), val = int32(0)];
+            int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)];
+            bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [849, 640]> decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_9")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)];
+            int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)];
+            string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_8")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")];
+            int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)];
+            int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)];
+            string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")];
+            bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)];
+            string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")];
+            string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1086848)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4363712)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7640576)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = string("transpose_4")];
+            tensor<fp16, [1, 1, 640]> input_5_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")];
+            string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")];
+            bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)];
+            string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")];
+            string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7645760)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10922624)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14199488)))];
+            tensor<fp16, [1, 1, 640]> input_5_cast_fp16_0, tensor<fp16, [1, 640]> input_5_cast_fp16_1, tensor<fp16, [1, 640]> input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")];
+            int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")];
+            string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")];
+            string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [3]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [3]> input_7_perm_0 = const()[name = string("input_7_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [640, 1024]> joint_module_enc_weight_to_fp16 = const()[name = string("joint_module_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14204672)))];
+            tensor<fp16, [640]> joint_module_enc_bias_to_fp16 = const()[name = string("joint_module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15515456)))];
+            tensor<fp16, [1, 1024, 1]> encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_4")];
+            tensor<fp16, [1, 1, 1024]> input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = encoder_to_fp16)[name = string("transpose_2")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15516800)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16336064)))];
+            tensor<fp16, [1, 1, 640]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_3")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_79_axes_0 = const()[name = string("op_79_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_79_cast_fp16 = expand_dims(axes = var_79_axes_0, x = linear_0_cast_fp16)[name = string("op_79_cast_fp16")];
+            tensor<int32, [1]> var_80_axes_0 = const()[name = string("op_80_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_80_cast_fp16 = expand_dims(axes = var_80_axes_0, x = linear_1_cast_fp16)[name = string("op_80_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_11_cast_fp16 = add(x = var_79_cast_fp16, y = var_80_cast_fp16)[name = string("input_11_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_13_cast_fp16 = relu(x = input_11_cast_fp16)[name = string("input_13_cast_fp16")];
+            tensor<fp16, [849, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16337408)))];
+            tensor<fp16, [849]> joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [849]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17424192)))];
+            tensor<fp16, [1, 1, 1, 849]> linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_2_cast_fp16")];
+            string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 1, 1, 849]> logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_3")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_5")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_6")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")];
+        } -> (logits, h_out, c_out);
+}
\ No newline at end of file
diff --git a/fr/1120ms/decoder_joint.mlmodelc/weights/weight.bin b/fr/1120ms/decoder_joint.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..7e387cabbe194b0f606c14bdf2b03b19d697a8f3
--- /dev/null
+++ b/fr/1120ms/decoder_joint.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39ce7eecdfb6a432f391c931412d9a5cd15fb0ef85562ef0c207e4a18fe6b50e
+size 17425954
diff --git a/fr/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..8de2f672a8b40625e6b0f13654d4cba788880b07
--- /dev/null
+++ b/fr/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72ce6705a73b8b0488f31dd35a385908473aff8bef0ffe290f56bb51068a1d26
+size 13745
diff --git a/fr/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..7e387cabbe194b0f606c14bdf2b03b19d697a8f3
--- /dev/null
+++ b/fr/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39ce7eecdfb6a432f391c931412d9a5cd15fb0ef85562ef0c207e4a18fe6b50e
+size 17425954
diff --git a/fr/1120ms/decoder_joint.mlpackage/Manifest.json b/fr/1120ms/decoder_joint.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..967aa622f9a5ac2f95978e9feda32d61c8a73643
--- /dev/null
+++ b/fr/1120ms/decoder_joint.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "8FCEBC80-AF4D-42E1-B70D-96BB47913BC7": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "A77788EC-CA36-4DF3-BB1B-E6182E0ABF28": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "8FCEBC80-AF4D-42E1-B70D-96BB47913BC7"
+}
diff --git a/fr/1120ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin b/fr/1120ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..425b9fbc36cf7602d11ebba8a2e29be972229238
--- /dev/null
+++ b/fr/1120ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:878825967d0c03e7b016222a7b824badb35d1d7e134ad3a973abc00dc5fe353a
+size 243
diff --git a/fr/1120ms/decoder_joint_noencproj.mlmodelc/coremldata.bin b/fr/1120ms/decoder_joint_noencproj.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..479d8bc5a3eab6af7208785446fe73a569a877dc
--- /dev/null
+++ b/fr/1120ms/decoder_joint_noencproj.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ebb7868572e2691055cf8d926fe61e2049bdc0a3e6349e82227f0b82dd849814
+size 519
diff --git a/fr/1120ms/decoder_joint_noencproj.mlmodelc/model.mil b/fr/1120ms/decoder_joint_noencproj.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..ef9fa8f205f75bf414b61d0b622d2976b441fe5b
--- /dev/null
+++ b/fr/1120ms/decoder_joint_noencproj.mlmodelc/model.mil
@@ -0,0 +1,91 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios18>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [1, 1, 640]> encoder_proj, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)];
+            bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [849, 640]> decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")];
+            string cast_1_dtype_0 = const()[name = string("cast_1_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_10")];
+            tensor<int32, [1, 1]> cast_1 = cast(dtype = cast_1_dtype_0, x = token_to_int16)[name = string("cast_9")];
+            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
+            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(849)];
+            tensor<int32, [1, 1]> add_2 = add(x = cast_1, y = slice_by_index_0)[name = string("add_2")];
+            tensor<int32, [1, 1]> select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = string("select_0")];
+            int32 y_cast_fp16_cast_uint16_axis_0 = const()[name = string("y_cast_fp16_cast_uint16_axis_0"), val = int32(0)];
+            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
+            tensor<int16, [1, 1]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_8")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)];
+            int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)];
+            string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")];
+            int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)];
+            int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)];
+            string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")];
+            bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)];
+            string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")];
+            string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1086848)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4363712)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7640576)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = string("transpose_3")];
+            tensor<fp16, [1, 1, 640]> input_5_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")];
+            string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")];
+            bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)];
+            string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")];
+            string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7645760)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10922624)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14199488)))];
+            tensor<fp16, [1, 1, 640]> input_5_cast_fp16_0, tensor<fp16, [1, 640]> input_5_cast_fp16_1, tensor<fp16, [1, 640]> input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")];
+            int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")];
+            string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")];
+            string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [3]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14204672)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15023936)))];
+            tensor<fp16, [1, 1, 640]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_2")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [1]> f_axes_0 = const()[name = string("f_axes_0"), val = tensor<int32, [1]>([2])];
+            string encoder_proj_to_fp16_dtype_0 = const()[name = string("encoder_proj_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 1, 640]> encoder_proj_to_fp16 = cast(dtype = encoder_proj_to_fp16_dtype_0, x = encoder_proj)[name = string("cast_3")];
+            tensor<fp16, [1, 1, 1, 640]> f_cast_fp16 = expand_dims(axes = f_axes_0, x = encoder_proj_to_fp16)[name = string("f_cast_fp16")];
+            tensor<int32, [1]> g_axes_0 = const()[name = string("g_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> g_cast_fp16 = expand_dims(axes = g_axes_0, x = linear_0_cast_fp16)[name = string("g_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_9_cast_fp16 = add(x = f_cast_fp16, y = g_cast_fp16)[name = string("input_9_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_11_cast_fp16 = relu(x = input_9_cast_fp16)[name = string("input_11_cast_fp16")];
+            tensor<fp16, [849, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15025280)))];
+            tensor<fp16, [849]> joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [849]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16112064)))];
+            tensor<fp16, [1, 1, 1, 849]> linear_1_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_1_cast_fp16")];
+            int32 var_83 = const()[name = string("op_83"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1, 849]> var_85_softmax_cast_fp16 = softmax(axis = var_83, x = linear_1_cast_fp16)[name = string("op_85_softmax_cast_fp16")];
+            fp32 var_85_epsilon_0 = const()[name = string("op_85_epsilon_0"), val = fp32(0x1p-149)];
+            tensor<fp16, [1, 1, 1, 849]> var_85_cast_fp16 = log(epsilon = var_85_epsilon_0, x = var_85_softmax_cast_fp16)[name = string("op_85_cast_fp16")];
+            string var_85_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_85_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 1, 1, 849]> logits = cast(dtype = var_85_cast_fp16_to_fp32_dtype_0, x = var_85_cast_fp16)[name = string("cast_2")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")];
+        } -> (logits, h_out, c_out);
+}
\ No newline at end of file
diff --git a/fr/1120ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin b/fr/1120ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..e37358fc0d446e079fe87526671dc411c4b37591
--- /dev/null
+++ b/fr/1120ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c691d3004f2ac6a956208569a29d899a7f8e5030087d53fbd488785fa0e8efad
+size 16113826
diff --git a/fr/1120ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/1120ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..b701505b2faf5efd4a3c1024077f4834b1f94d37
--- /dev/null
+++ b/fr/1120ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4196a821e716db75328fdd2f99f83214ecd9e6e6481a96a88907fe1c4232f143
+size 14630
diff --git a/fr/1120ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/1120ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..e37358fc0d446e079fe87526671dc411c4b37591
--- /dev/null
+++ b/fr/1120ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c691d3004f2ac6a956208569a29d899a7f8e5030087d53fbd488785fa0e8efad
+size 16113826
diff --git a/fr/1120ms/decoder_joint_noencproj.mlpackage/Manifest.json b/fr/1120ms/decoder_joint_noencproj.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..4709121d3572643a899871aa076a46f4c02a7ef8
--- /dev/null
+++ b/fr/1120ms/decoder_joint_noencproj.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "1D4938C1-F035-4D52-9376-C0FB72335508": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "BF28BA0A-268B-4C46-AB97-C0173D8B55F0": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "BF28BA0A-268B-4C46-AB97-C0173D8B55F0"
+}
diff --git a/fr/1120ms/encoder.mlmodelc/analytics/coremldata.bin b/fr/1120ms/encoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..cdc6c00cf855a2b827e4760c156a421699a9f39e
--- /dev/null
+++ b/fr/1120ms/encoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0638780288ed25026170632a153dbed5d798954deac843fa58ddaae3190914d4
+size 243
diff --git a/fr/1120ms/encoder.mlmodelc/coremldata.bin b/fr/1120ms/encoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..10b9a6140d4cd343ceb8fc17695d32a4932d229d
--- /dev/null
+++ b/fr/1120ms/encoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5df783ab0c78ea6386cb181e85e1458a9cc1a8f5e06a49833b7b7a4b017105ed
+size 662
diff --git a/fr/1120ms/encoder.mlmodelc/model.mil b/fr/1120ms/encoder.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..8243eb3dcf4d99803bf1e78f45f94ceda08ed8ee
--- /dev/null
+++ b/fr/1120ms/encoder.mlmodelc/model.mil
@@ -0,0 +1,4439 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})]
+{
+    func main<ios18>(tensor<fp32, [1, 24, 42, 1024]> cache_channel, tensor<int32, [1]> cache_len, tensor<fp32, [1, 24, 1024, 8]> cache_time, tensor<fp32, [1, 128, 121]> mel, tensor<int32, [1]> mel_length, tensor<int32, [1]> prompt_id) {
+            tensor<int32, [4]> value_3_perm_0 = const()[name = string("value_3_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string cache_channel_to_fp16_dtype_0 = const()[name = string("cache_channel_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<int32, [4]> value_5_perm_0 = const()[name = string("value_5_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string cache_time_to_fp16_dtype_0 = const()[name = string("cache_time_to_fp16_dtype_0"), val = string("fp16")];
+            int32 var_60 = const()[name = string("op_60"), val = int32(-1)];
+            int32 var_69 = const()[name = string("op_69"), val = int32(1)];
+            tensor<int32, [3]> x_1_perm_0 = const()[name = string("x_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string mel_to_fp16_dtype_0 = const()[name = string("mel_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<int32, [1]> tensor_1_axes_0 = const()[name = string("tensor_1_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 128, 121]> mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = string("cast_22")];
+            tensor<fp16, [1, 121, 128]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = mel_to_fp16)[name = string("transpose_367")];
+            tensor<fp16, [1, 1, 121, 128]> tensor_1_cast_fp16 = expand_dims(axes = tensor_1_axes_0, x = x_1_cast_fp16)[name = string("tensor_1_cast_fp16")];
+            tensor<int32, [1, 121]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1, 121]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<int32, [1]> var_138_axes_0 = const()[name = string("op_138_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_138 = expand_dims(axes = var_138_axes_0, x = mel_length)[name = string("op_138")];
+            tensor<bool, [1, 121]> time_mask_1 = less(x = expand_dims_0, y = var_138)[name = string("time_mask_1")];
+            tensor<int32, [1]> var_140_axes_0 = const()[name = string("op_140_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 121, 1]> var_140 = expand_dims(axes = var_140_axes_0, x = time_mask_1)[name = string("op_140")];
+            tensor<int32, [3]> var_142_reps_0 = const()[name = string("op_142_reps_0"), val = tensor<int32, [3]>([1, 1, 128])];
+            tensor<bool, [1, 121, 128]> var_142 = tile(reps = var_142_reps_0, x = var_140)[name = string("op_142")];
+            tensor<int32, [1]> var_148_axes_0 = const()[name = string("op_148_axes_0"), val = tensor<int32, [1]>([1])];
+            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 121, 128]> var_142_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_142)[name = string("cast_21")];
+            tensor<fp16, [1, 1, 121, 128]> var_148_cast_fp16 = expand_dims(axes = var_148_axes_0, x = var_142_to_fp16)[name = string("op_148_cast_fp16")];
+            tensor<fp16, [1, 1, 121, 128]> input_1_cast_fp16 = mul(x = tensor_1_cast_fp16, y = var_148_cast_fp16)[name = string("input_1_cast_fp16")];
+            tensor<int32, [8]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("constant")];
+            fp16 const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 124, 131]> input_3_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")];
+            string tensor_3_pad_type_0 = const()[name = string("tensor_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_3_strides_0 = const()[name = string("tensor_3_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, [4]> tensor_3_pad_0 = const()[name = string("tensor_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_3_dilations_0 = const()[name = string("tensor_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 tensor_3_groups_0 = const()[name = string("tensor_3_groups_0"), val = int32(1)];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_0_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3008))))[name = string("encoder_pre_encode_conv_0_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3584)))];
+            tensor<fp16, [1, 256, 61, 65]> tensor_3_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = tensor_3_dilations_0, groups = tensor_3_groups_0, pad = tensor_3_pad_0, pad_type = tensor_3_pad_type_0, strides = tensor_3_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("tensor_3_cast_fp16")];
+            string current_lengths_1_to_fp16_dtype_0 = const()[name = string("current_lengths_1_to_fp16_dtype_0"), val = string("fp16")];
+            fp16 var_161_promoted_to_fp16 = const()[name = string("op_161_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> mel_length_to_fp16 = cast(dtype = current_lengths_1_to_fp16_dtype_0, x = mel_length)[name = string("cast_20")];
+            tensor<fp16, [1]> var_162_cast_fp16 = add(x = mel_length_to_fp16, y = var_161_promoted_to_fp16)[name = string("op_162_cast_fp16")];
+            fp16 var_163_promoted_to_fp16 = const()[name = string("op_163_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> var_164_cast_fp16 = add(x = var_162_cast_fp16, y = var_163_promoted_to_fp16)[name = string("op_164_cast_fp16")];
+            fp16 var_165_promoted_to_fp16 = const()[name = string("op_165_promoted_to_fp16"), val = fp16(0x1.8p+1)];
+            tensor<fp16, [1]> var_166_cast_fp16 = sub(x = var_164_cast_fp16, y = var_165_promoted_to_fp16)[name = string("op_166_cast_fp16")];
+            fp16 var_57_promoted_to_fp16 = const()[name = string("op_57_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> floor_div_0_cast_fp16 = floor_div(x = var_166_cast_fp16, y = var_57_promoted_to_fp16)[name = string("floor_div_0_cast_fp16")];
+            fp16 var_168_promoted_to_fp16 = const()[name = string("op_168_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_3_cast_fp16 = add(x = floor_div_0_cast_fp16, y = var_168_promoted_to_fp16)[name = string("current_lengths_3_cast_fp16")];
+            string lengths_19_dtype_0 = const()[name = string("lengths_19_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 61]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1, 61]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4160)))];
+            tensor<int32, [1]> var_177_axes_0 = const()[name = string("op_177_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_3_cast_fp16_to_int32 = cast(dtype = lengths_19_dtype_0, x = current_lengths_3_cast_fp16)[name = string("cast_19")];
+            tensor<int32, [1, 1]> var_177 = expand_dims(axes = var_177_axes_0, x = current_lengths_3_cast_fp16_to_int32)[name = string("op_177")];
+            tensor<bool, [1, 61]> time_mask_3 = less(x = expand_dims_1, y = var_177)[name = string("time_mask_3")];
+            tensor<int32, [1]> var_179_axes_0 = const()[name = string("op_179_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 61, 1]> var_179 = expand_dims(axes = var_179_axes_0, x = time_mask_3)[name = string("op_179")];
+            tensor<int32, [3]> var_181_reps_0 = const()[name = string("op_181_reps_0"), val = tensor<int32, [3]>([1, 1, 65])];
+            tensor<bool, [1, 61, 65]> var_181 = tile(reps = var_181_reps_0, x = var_179)[name = string("op_181")];
+            tensor<int32, [1]> var_187_axes_0 = const()[name = string("op_187_axes_0"), val = tensor<int32, [1]>([1])];
+            string mask_3_to_fp16_dtype_0 = const()[name = string("mask_3_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 61, 65]> var_181_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_181)[name = string("cast_18")];
+            tensor<fp16, [1, 1, 61, 65]> var_187_cast_fp16 = expand_dims(axes = var_187_axes_0, x = var_181_to_fp16)[name = string("op_187_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_3_reps_0 = const()[name = string("expanded_mask_3_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 61, 65]> expanded_mask_3_cast_fp16 = tile(reps = expanded_mask_3_reps_0, x = var_187_cast_fp16)[name = string("expanded_mask_3_cast_fp16")];
+            tensor<fp16, [1, 256, 61, 65]> input_5_cast_fp16 = mul(x = tensor_3_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<fp16, [1, 256, 61, 65]> tensor_5_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("tensor_5_cast_fp16")];
+            tensor<fp16, [1, 256, 61, 65]> input_7_cast_fp16 = mul(x = tensor_5_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_7_cast_fp16")];
+            tensor<int32, [8]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            string input_9_mode_0 = const()[name = string("input_9_mode_0"), val = string("constant")];
+            fp16 const_23_to_fp16 = const()[name = string("const_23_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 256, 64, 68]> input_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")];
+            string tensor_7_pad_type_0 = const()[name = string("tensor_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_7_strides_0 = const()[name = string("tensor_7_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            int32 tensor_7_groups_0 = const()[name = string("tensor_7_groups_0"), val = int32(256)];
+            tensor<int32, [4]> tensor_7_pad_0 = const()[name = string("tensor_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_7_dilations_0 = const()[name = string("tensor_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4480))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6848))))[name = string("encoder_pre_encode_conv_2_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7424)))];
+            tensor<fp16, [1, 256, 31, 33]> tensor_7_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = tensor_7_dilations_0, groups = tensor_7_groups_0, pad = tensor_7_pad_0, pad_type = tensor_7_pad_type_0, strides = tensor_7_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("tensor_7_cast_fp16")];
+            fp16 var_209_promoted_to_fp16 = const()[name = string("op_209_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> var_210_cast_fp16 = add(x = current_lengths_3_cast_fp16, y = var_209_promoted_to_fp16)[name = string("op_210_cast_fp16")];
+            fp16 var_211_promoted_to_fp16 = const()[name = string("op_211_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> var_212_cast_fp16 = add(x = var_210_cast_fp16, y = var_211_promoted_to_fp16)[name = string("op_212_cast_fp16")];
+            fp16 var_213_promoted_to_fp16 = const()[name = string("op_213_promoted_to_fp16"), val = fp16(0x1.8p+1)];
+            tensor<fp16, [1]> var_214_cast_fp16 = sub(x = var_212_cast_fp16, y = var_213_promoted_to_fp16)[name = string("op_214_cast_fp16")];
+            fp16 var_57_promoted_1_to_fp16 = const()[name = string("op_57_promoted_1_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> floor_div_1_cast_fp16 = floor_div(x = var_214_cast_fp16, y = var_57_promoted_1_to_fp16)[name = string("floor_div_1_cast_fp16")];
+            fp16 var_216_promoted_to_fp16 = const()[name = string("op_216_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_5_cast_fp16 = add(x = floor_div_1_cast_fp16, y = var_216_promoted_to_fp16)[name = string("current_lengths_5_cast_fp16")];
+            string lengths_21_dtype_0 = const()[name = string("lengths_21_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 31]> expand_dims_2 = const()[name = string("expand_dims_2"), val = tensor<int32, [1, 31]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8000)))];
+            tensor<int32, [1]> var_225_axes_0 = const()[name = string("op_225_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_5_cast_fp16_to_int32 = cast(dtype = lengths_21_dtype_0, x = current_lengths_5_cast_fp16)[name = string("cast_17")];
+            tensor<int32, [1, 1]> var_225 = expand_dims(axes = var_225_axes_0, x = current_lengths_5_cast_fp16_to_int32)[name = string("op_225")];
+            tensor<bool, [1, 31]> time_mask_5 = less(x = expand_dims_2, y = var_225)[name = string("time_mask_5")];
+            tensor<int32, [1]> var_227_axes_0 = const()[name = string("op_227_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 31, 1]> var_227 = expand_dims(axes = var_227_axes_0, x = time_mask_5)[name = string("op_227")];
+            tensor<int32, [3]> var_229_reps_0 = const()[name = string("op_229_reps_0"), val = tensor<int32, [3]>([1, 1, 33])];
+            tensor<bool, [1, 31, 33]> var_229 = tile(reps = var_229_reps_0, x = var_227)[name = string("op_229")];
+            tensor<int32, [1]> var_235_axes_0 = const()[name = string("op_235_axes_0"), val = tensor<int32, [1]>([1])];
+            string mask_5_to_fp16_dtype_0 = const()[name = string("mask_5_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 31, 33]> var_229_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_229)[name = string("cast_16")];
+            tensor<fp16, [1, 1, 31, 33]> var_235_cast_fp16 = expand_dims(axes = var_235_axes_0, x = var_229_to_fp16)[name = string("op_235_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_7_reps_0 = const()[name = string("expanded_mask_7_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 31, 33]> expanded_mask_7_cast_fp16 = tile(reps = expanded_mask_7_reps_0, x = var_235_cast_fp16)[name = string("expanded_mask_7_cast_fp16")];
+            tensor<fp16, [1, 256, 31, 33]> input_11_cast_fp16 = mul(x = tensor_7_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_11_cast_fp16")];
+            string tensor_9_pad_type_0 = const()[name = string("tensor_9_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_9_strides_0 = const()[name = string("tensor_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> tensor_9_pad_0 = const()[name = string("tensor_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_9_dilations_0 = const()[name = string("tensor_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 tensor_9_groups_0 = const()[name = string("tensor_9_groups_0"), val = int32(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_3_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8192))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73792))))[name = string("encoder_pre_encode_conv_3_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74368)))];
+            tensor<fp16, [1, 256, 31, 33]> tensor_9_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = tensor_9_dilations_0, groups = tensor_9_groups_0, pad = tensor_9_pad_0, pad_type = tensor_9_pad_type_0, strides = tensor_9_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("tensor_9_cast_fp16")];
+            tensor<fp16, [1, 256, 31, 33]> input_13_cast_fp16 = mul(x = tensor_9_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_13_cast_fp16")];
+            tensor<fp16, [1, 256, 31, 33]> tensor_11_cast_fp16 = relu(x = input_13_cast_fp16)[name = string("tensor_11_cast_fp16")];
+            tensor<fp16, [1, 256, 31, 33]> input_15_cast_fp16 = mul(x = tensor_11_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_15_cast_fp16")];
+            tensor<int32, [8]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            string input_17_mode_0 = const()[name = string("input_17_mode_0"), val = string("constant")];
+            fp16 const_41_to_fp16 = const()[name = string("const_41_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 256, 34, 36]> input_17_cast_fp16 = pad(constant_val = const_41_to_fp16, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")];
+            string tensor_13_pad_type_0 = const()[name = string("tensor_13_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_13_strides_0 = const()[name = string("tensor_13_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            int32 tensor_13_groups_0 = const()[name = string("tensor_13_groups_0"), val = int32(256)];
+            tensor<int32, [4]> tensor_13_pad_0 = const()[name = string("tensor_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_13_dilations_0 = const()[name = string("tensor_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_5_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74944))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77312))))[name = string("encoder_pre_encode_conv_5_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77888)))];
+            tensor<fp16, [1, 256, 16, 17]> tensor_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = tensor_13_dilations_0, groups = tensor_13_groups_0, pad = tensor_13_pad_0, pad_type = tensor_13_pad_type_0, strides = tensor_13_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("tensor_13_cast_fp16")];
+            fp16 var_272_promoted_to_fp16 = const()[name = string("op_272_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> var_273_cast_fp16 = add(x = current_lengths_5_cast_fp16, y = var_272_promoted_to_fp16)[name = string("op_273_cast_fp16")];
+            fp16 var_274_promoted_to_fp16 = const()[name = string("op_274_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> var_275_cast_fp16 = add(x = var_273_cast_fp16, y = var_274_promoted_to_fp16)[name = string("op_275_cast_fp16")];
+            fp16 var_276_promoted_to_fp16 = const()[name = string("op_276_promoted_to_fp16"), val = fp16(0x1.8p+1)];
+            tensor<fp16, [1]> var_277_cast_fp16 = sub(x = var_275_cast_fp16, y = var_276_promoted_to_fp16)[name = string("op_277_cast_fp16")];
+            fp16 var_57_promoted_2_to_fp16 = const()[name = string("op_57_promoted_2_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> floor_div_2_cast_fp16 = floor_div(x = var_277_cast_fp16, y = var_57_promoted_2_to_fp16)[name = string("floor_div_2_cast_fp16")];
+            fp16 var_279_promoted_to_fp16 = const()[name = string("op_279_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_cast_fp16 = add(x = floor_div_2_cast_fp16, y = var_279_promoted_to_fp16)[name = string("current_lengths_cast_fp16")];
+            string lengths_dtype_0 = const()[name = string("lengths_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 16]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78464)))];
+            tensor<int32, [1]> var_288_axes_0 = const()[name = string("op_288_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_cast_fp16_to_int32 = cast(dtype = lengths_dtype_0, x = current_lengths_cast_fp16)[name = string("cast_15")];
+            tensor<int32, [1, 1]> var_288 = expand_dims(axes = var_288_axes_0, x = current_lengths_cast_fp16_to_int32)[name = string("op_288")];
+            tensor<bool, [1, 16]> time_mask = less(x = expand_dims_3, y = var_288)[name = string("time_mask")];
+            tensor<int32, [1]> var_290_axes_0 = const()[name = string("op_290_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 16, 1]> var_290 = expand_dims(axes = var_290_axes_0, x = time_mask)[name = string("op_290")];
+            tensor<int32, [3]> var_292_reps_0 = const()[name = string("op_292_reps_0"), val = tensor<int32, [3]>([1, 1, 17])];
+            tensor<bool, [1, 16, 17]> var_292 = tile(reps = var_292_reps_0, x = var_290)[name = string("op_292")];
+            tensor<int32, [1]> var_298_axes_0 = const()[name = string("op_298_axes_0"), val = tensor<int32, [1]>([1])];
+            string mask_7_to_fp16_dtype_0 = const()[name = string("mask_7_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 16, 17]> var_292_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_292)[name = string("cast_14")];
+            tensor<fp16, [1, 1, 16, 17]> var_298_cast_fp16 = expand_dims(axes = var_298_axes_0, x = var_292_to_fp16)[name = string("op_298_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_13_reps_0 = const()[name = string("expanded_mask_13_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 16, 17]> expanded_mask_13_cast_fp16 = tile(reps = expanded_mask_13_reps_0, x = var_298_cast_fp16)[name = string("expanded_mask_13_cast_fp16")];
+            tensor<fp16, [1, 256, 16, 17]> input_19_cast_fp16 = mul(x = tensor_13_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_19_cast_fp16")];
+            string tensor_15_pad_type_0 = const()[name = string("tensor_15_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_15_strides_0 = const()[name = string("tensor_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> tensor_15_pad_0 = const()[name = string("tensor_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_15_dilations_0 = const()[name = string("tensor_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 tensor_15_groups_0 = const()[name = string("tensor_15_groups_0"), val = int32(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_6_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78592))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144192))))[name = string("encoder_pre_encode_conv_6_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144768)))];
+            tensor<fp16, [1, 256, 16, 17]> tensor_15_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = tensor_15_dilations_0, groups = tensor_15_groups_0, pad = tensor_15_pad_0, pad_type = tensor_15_pad_type_0, strides = tensor_15_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("tensor_15_cast_fp16")];
+            tensor<fp16, [1, 256, 16, 17]> input_21_cast_fp16 = mul(x = tensor_15_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_21_cast_fp16")];
+            tensor<fp16, [1, 256, 16, 17]> tensor_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("tensor_cast_fp16")];
+            tensor<fp16, [1, 256, 16, 17]> x_3_cast_fp16 = mul(x = tensor_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("x_3_cast_fp16")];
+            tensor<int32, [4]> var_332_perm_0 = const()[name = string("op_332_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_333 = const()[name = string("op_333"), val = tensor<int32, [3]>([1, 16, -1])];
+            tensor<fp16, [1, 16, 256, 17]> var_332_cast_fp16 = transpose(perm = var_332_perm_0, x = x_3_cast_fp16)[name = string("transpose_366")];
+            tensor<fp16, [1, 16, 4352]> input_23_cast_fp16 = reshape(shape = var_333, x = var_332_cast_fp16)[name = string("input_23_cast_fp16")];
+            tensor<fp16, [1024, 4352]> encoder_pre_encode_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4352]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145344))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4601856))))[name = string("encoder_pre_encode_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_pre_encode_out_bias_to_fp16 = const()[name = string("encoder_pre_encode_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4603968)))];
+            tensor<fp16, [1, 16, 1024]> linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [3]> var_343_begin_0 = const()[name = string("op_343_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_343_end_0 = const()[name = string("op_343_end_0"), val = tensor<int32, [3]>([1, 16, 1024])];
+            tensor<bool, [3]> var_343_end_mask_0 = const()[name = string("op_343_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_343_cast_fp16 = slice_by_index(begin = var_343_begin_0, end = var_343_end_0, end_mask = var_343_end_mask_0, x = linear_0_cast_fp16)[name = string("op_343_cast_fp16")];
+            int32 var_345 = const()[name = string("op_345"), val = int32(2)];
+            tensor<int32, [1]> var_346 = sub(x = current_lengths_cast_fp16_to_int32, y = var_345)[name = string("op_346")];
+            string var_346_promoted_to_fp16_dtype_0 = const()[name = string("op_346_promoted_to_fp16_dtype_0"), val = string("fp16")];
+            fp16 var_63_promoted_to_fp16 = const()[name = string("op_63_promoted_to_fp16"), val = fp16(0x0p+0)];
+            fp16 const_61_to_fp16 = const()[name = string("const_61_to_fp16"), val = fp16(inf)];
+            tensor<fp16, [1]> var_346_to_fp16 = cast(dtype = var_346_promoted_to_fp16_dtype_0, x = var_346)[name = string("cast_13")];
+            tensor<fp16, [1]> clip_0_cast_fp16 = clip(alpha = var_63_promoted_to_fp16, beta = const_61_to_fp16, x = var_346_to_fp16)[name = string("clip_0_cast_fp16")];
+            tensor<int32, [1]> max_audio_length_1 = const()[name = string("max_audio_length_1"), val = tensor<int32, [1]>([14])];
+            fp16 var_362_promoted_to_fp16 = const()[name = string("op_362_promoted_to_fp16"), val = fp16(0x1.5p+5)];
+            tensor<fp16, [1]> padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_362_promoted_to_fp16)[name = string("padding_length_cast_fp16")];
+            int32 const_63 = const()[name = string("const_63"), val = int32(-1)];
+            tensor<int32, [1]> var_364 = mul(x = cache_len, y = const_63)[name = string("op_364")];
+            int32 var_365 = const()[name = string("op_365"), val = int32(42)];
+            tensor<int32, [1]> offset = add(x = var_364, y = var_365)[name = string("offset")];
+            tensor<int32, [1]> var_405_axes_0 = const()[name = string("op_405_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 1]> var_405_cast_fp16 = expand_dims(axes = var_405_axes_0, x = padding_length_cast_fp16)[name = string("op_405_cast_fp16")];
+            tensor<fp16, [1, 56]> var_404_promoted_to_fp16 = const()[name = string("op_404_promoted_to_fp16"), val = tensor<fp16, [1, 56]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606080)))];
+            tensor<bool, [1, 56]> pad_mask_1_cast_fp16 = less(x = var_404_promoted_to_fp16, y = var_405_cast_fp16)[name = string("pad_mask_1_cast_fp16")];
+            tensor<int32, [1, 56]> expand_dims_5 = const()[name = string("expand_dims_5"), val = tensor<int32, [1, 56]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606272)))];
+            tensor<int32, [1]> var_411_axes_0 = const()[name = string("op_411_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 1]> var_411 = expand_dims(axes = var_411_axes_0, x = offset)[name = string("op_411")];
+            tensor<bool, [1, 56]> pad_mask_off = greater_equal(x = expand_dims_5, y = var_411)[name = string("pad_mask_off")];
+            tensor<bool, [1, 56]> pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = string("pad_mask_3")];
+            tensor<int32, [1]> var_414_axes_0 = const()[name = string("op_414_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 56]> var_414 = expand_dims(axes = var_414_axes_0, x = pad_mask_3)[name = string("op_414")];
+            tensor<int32, [3]> var_415 = const()[name = string("op_415"), val = tensor<int32, [3]>([1, 56, 1])];
+            tensor<bool, [1, 56, 56]> pad_mask_for_att_mask_1 = tile(reps = var_415, x = var_414)[name = string("pad_mask_for_att_mask_1")];
+            tensor<int32, [3]> var_417_perm_0 = const()[name = string("op_417_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<bool, [1, 56, 56]> var_417 = transpose(perm = var_417_perm_0, x = pad_mask_for_att_mask_1)[name = string("transpose_365")];
+            tensor<bool, [1, 56, 56]> pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_417)[name = string("pad_mask_for_att_mask")];
+            tensor<bool, [1, 56, 56]> const_71 = const()[name = string("const_71"), val = tensor<bool, [1, 56, 56]>([[[true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])];
+            tensor<bool, [1, 56, 56]> att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_71)[name = string("att_mask_9")];
+            tensor<bool, [1, 56, 56]> att_mask = logical_not(x = att_mask_9)[name = string("att_mask")];
+            tensor<bool, [1, 56]> pad_mask_5 = logical_not(x = pad_mask_3)[name = string("pad_mask_5")];
+            tensor<int32, [2]> pad_mask_begin_0 = const()[name = string("pad_mask_begin_0"), val = tensor<int32, [2]>([0, 42])];
+            tensor<int32, [2]> pad_mask_end_0 = const()[name = string("pad_mask_end_0"), val = tensor<int32, [2]>([1, 56])];
+            tensor<bool, [2]> pad_mask_end_mask_0 = const()[name = string("pad_mask_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<bool, [1, 14]> pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = string("pad_mask")];
+            tensor<int32, [3]> mask_9_begin_0 = const()[name = string("mask_9_begin_0"), val = tensor<int32, [3]>([0, 42, 0])];
+            tensor<int32, [3]> mask_9_end_0 = const()[name = string("mask_9_end_0"), val = tensor<int32, [3]>([1, 56, 56])];
+            tensor<bool, [3]> mask_9_end_mask_0 = const()[name = string("mask_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<bool, [1, 14, 56]> mask_9 = slice_by_index(begin = mask_9_begin_0, end = mask_9_end_0, end_mask = mask_9_end_mask_0, x = att_mask)[name = string("mask_9")];
+            tensor<int32, [4]> cache_1_begin_0 = const()[name = string("cache_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_1_end_0 = const()[name = string("cache_1_end_0"), val = tensor<int32, [4]>([1, 1, 42, 1024])];
+            tensor<bool, [4]> cache_1_end_mask_0 = const()[name = string("cache_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_1_squeeze_mask_0 = const()[name = string("cache_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 24, 42, 1024]> cache_channel_to_fp16 = cast(dtype = cache_channel_to_fp16_dtype_0, x = cache_channel)[name = string("cast_12")];
+            tensor<fp16, [24, 1, 42, 1024]> value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = cache_channel_to_fp16)[name = string("transpose_364")];
+            tensor<fp16, [1, 42, 1024]> cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_1_cast_fp16")];
+            tensor<int32, [4]> cache_3_begin_0 = const()[name = string("cache_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_3_end_0 = const()[name = string("cache_3_end_0"), val = tensor<int32, [4]>([1, 1, 1024, 8])];
+            tensor<bool, [4]> cache_3_end_mask_0 = const()[name = string("cache_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_3_squeeze_mask_0 = const()[name = string("cache_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 24, 1024, 8]> cache_time_to_fp16 = cast(dtype = cache_time_to_fp16_dtype_0, x = cache_time)[name = string("cast_11")];
+            tensor<fp16, [24, 1, 1024, 8]> value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = cache_time_to_fp16)[name = string("transpose_363")];
+            tensor<fp16, [1, 1024, 8]> cache_3_cast_fp16 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_3_cast_fp16")];
+            tensor<int32, [1]> input_27_axes_0 = const()[name = string("input_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606592)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4608704)))];
+            fp16 var_43_to_fp16 = const()[name = string("op_43_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 14, 1024]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_343_cast_fp16)[name = string("input_27_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4610816))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8805184))))[name = string("encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_0_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8813440)))];
+            tensor<fp16, [1, 14, 4096]> linear_1_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = string("input_31_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8821696))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13016064))))[name = string("encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13018176)))];
+            tensor<fp16, [1, 14, 1024]> linear_2_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_2_cast_fp16")];
+            fp16 var_456_to_fp16 = const()[name = string("op_456_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_457_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_456_to_fp16)[name = string("op_457_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_37_cast_fp16 = add(x = var_343_cast_fp16, y = var_457_cast_fp16)[name = string("input_37_cast_fp16")];
+            tensor<int32, [1]> key_1_axes_0 = const()[name = string("key_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13020288)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13022400)))];
+            tensor<fp16, [1, 14, 1024]> key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = string("key_1_cast_fp16")];
+            bool input_39_interleave_0 = const()[name = string("input_39_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_39_cast_fp16 = concat(axis = var_69, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = string("input_39_cast_fp16")];
+            tensor<int32, [3]> var_479_begin_0 = const()[name = string("op_479_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_479_end_0 = const()[name = string("op_479_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_479_end_mask_0 = const()[name = string("op_479_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_479_cast_fp16 = slice_by_index(begin = var_479_begin_0, end = var_479_end_0, end_mask = var_479_end_mask_0, x = cache_1_cast_fp16)[name = string("op_479_cast_fp16")];
+            bool var_485_interleave_0 = const()[name = string("op_485_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_485_cast_fp16 = concat(axis = var_69, interleave = var_485_interleave_0, values = (var_479_cast_fp16, key_1_cast_fp16))[name = string("op_485_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13024512))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14073152))))[name = string("encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14075264)))];
+            tensor<fp16, [1, 14, 1024]> linear_3_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized, x = key_1_cast_fp16)[name = string("linear_3_cast_fp16")];
+            tensor<int32, [4]> var_490 = const()[name = string("op_490"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_1_cast_fp16 = reshape(shape = var_490, x = linear_3_cast_fp16)[name = string("q_1_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14077376))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15126016))))[name = string("encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15128128)))];
+            tensor<fp16, [1, 56, 1024]> linear_4_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_4_cast_fp16")];
+            tensor<int32, [4]> var_495 = const()[name = string("op_495"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_1_cast_fp16 = reshape(shape = var_495, x = linear_4_cast_fp16)[name = string("k_1_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15130240))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16178880))))[name = string("encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16180992)))];
+            tensor<fp16, [1, 56, 1024]> linear_5_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_5_cast_fp16")];
+            tensor<int32, [4]> var_500 = const()[name = string("op_500"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_1_cast_fp16 = reshape(shape = var_500, x = linear_5_cast_fp16)[name = string("v_1_cast_fp16")];
+            tensor<int32, [4]> value_9_perm_0 = const()[name = string("value_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16183104)))];
+            tensor<fp16, [1, 14, 8, 128]> var_513_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = string("op_513_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16185216)))];
+            tensor<fp16, [1, 14, 8, 128]> var_515_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = string("op_515_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_1_perm_0 = const()[name = string("q_with_bias_v_1_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_7_transpose_x_0 = const()[name = string("x_7_transpose_x_0"), val = bool(false)];
+            bool x_7_transpose_y_0 = const()[name = string("x_7_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_517_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16187328))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16301056))))[name = string("op_517_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_515_cast_fp16)[name = string("transpose_362")];
+            tensor<fp16, [1, 8, 14, 111]> x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = op_517_to_fp16_quantized)[name = string("x_7_cast_fp16")];
+            tensor<int32, [8]> x_9_pad_0 = const()[name = string("x_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_9_mode_0 = const()[name = string("x_9_mode_0"), val = string("constant")];
+            fp16 const_79_to_fp16 = const()[name = string("const_79_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_9_cast_fp16 = pad(constant_val = const_79_to_fp16, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7_cast_fp16)[name = string("x_9_cast_fp16")];
+            tensor<int32, [4]> var_525 = const()[name = string("op_525"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_11_cast_fp16 = reshape(shape = var_525, x = x_9_cast_fp16)[name = string("x_11_cast_fp16")];
+            tensor<int32, [4]> var_529_begin_0 = const()[name = string("op_529_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_529_end_0 = const()[name = string("op_529_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_529_end_mask_0 = const()[name = string("op_529_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_529_cast_fp16 = slice_by_index(begin = var_529_begin_0, end = var_529_end_0, end_mask = var_529_end_mask_0, x = x_11_cast_fp16)[name = string("op_529_cast_fp16")];
+            tensor<int32, [4]> var_530 = const()[name = string("op_530"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_1_cast_fp16 = reshape(shape = var_530, x = var_529_cast_fp16)[name = string("matrix_bd_1_cast_fp16")];
+            bool matrix_ac_1_transpose_x_0 = const()[name = string("matrix_ac_1_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_1_transpose_y_0 = const()[name = string("matrix_ac_1_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_97 = transpose(perm = transpose_97_perm_0, x = k_1_cast_fp16)[name = string("transpose_360")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_96 = transpose(perm = transpose_96_perm_0, x = var_513_cast_fp16)[name = string("transpose_361")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_96, y = transpose_97)[name = string("matrix_ac_1_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_3_begin_0 = const()[name = string("matrix_bd_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_3_end_0 = const()[name = string("matrix_bd_3_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_3_end_mask_0 = const()[name = string("matrix_bd_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1_cast_fp16)[name = string("matrix_bd_3_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_539_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = string("op_539_cast_fp16")];
+            fp16 _inversed_scores_1_y_0_to_fp16 = const()[name = string("_inversed_scores_1_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_1_cast_fp16 = mul(x = var_539_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = string("_inversed_scores_1_cast_fp16")];
+            tensor<int32, [1]> mask_11_axes_0 = const()[name = string("mask_11_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 14, 56]> mask_11 = expand_dims(axes = mask_11_axes_0, x = mask_9)[name = string("mask_11")];
+            fp16 var_46_to_fp16 = const()[name = string("op_46_to_fp16"), val = fp16(-0x1.388p+13)];
+            tensor<fp16, [1, 8, 14, 56]> scores_3_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_11)[name = string("scores_3_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_545_cast_fp16 = softmax(axis = var_60, x = scores_3_cast_fp16)[name = string("op_545_cast_fp16")];
+            fp16 var_45_to_fp16 = const()[name = string("op_45_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 56]> input_41_cast_fp16 = select(a = var_45_to_fp16, b = var_545_cast_fp16, cond = mask_11)[name = string("input_41_cast_fp16")];
+            bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)];
+            bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_1_cast_fp16)[name = string("transpose_359")];
+            tensor<fp16, [1, 8, 14, 128]> x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41_cast_fp16, y = value_9_cast_fp16)[name = string("x_13_cast_fp16")];
+            tensor<int32, [4]> var_549_perm_0 = const()[name = string("op_549_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_550 = const()[name = string("op_550"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_549_cast_fp16 = transpose(perm = var_549_perm_0, x = x_13_cast_fp16)[name = string("transpose_358")];
+            tensor<fp16, [1, 14, 1024]> input_43_cast_fp16 = reshape(shape = var_550, x = var_549_cast_fp16)[name = string("input_43_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16301376))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17350016))))[name = string("encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17352128)))];
+            tensor<fp16, [1, 14, 1024]> linear_7_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_7_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = string("input_47_cast_fp16")];
+            tensor<int32, [1]> x_17_axes_0 = const()[name = string("x_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17354240)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17356352)))];
+            tensor<fp16, [1, 14, 1024]> x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = string("x_17_cast_fp16")];
+            tensor<int32, [3]> input_49_perm_0 = const()[name = string("input_49_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_51_pad_type_0 = const()[name = string("input_51_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_51_strides_0 = const()[name = string("input_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_51_pad_0 = const()[name = string("input_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_51_dilations_0 = const()[name = string("input_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_51_groups_0 = const()[name = string("input_51_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17358464))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19455680))))[name = string("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = string("transpose_357")];
+            tensor<fp16, [1, 2048, 14]> input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("input_51_cast_fp16")];
+            int32 x_19_split_num_splits_0 = const()[name = string("x_19_split_num_splits_0"), val = int32(2)];
+            int32 x_19_split_axis_0 = const()[name = string("x_19_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_19_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51_cast_fp16)[name = string("x_19_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = string("x_19_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = string("x_19_cast_fp16")];
+            tensor<int32, [1]> var_576_axes_0 = const()[name = string("op_576_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 14]> var_576 = expand_dims(axes = var_576_axes_0, x = pad_mask)[name = string("op_576")];
+            tensor<fp16, [1, 1024, 14]> input_53_cast_fp16 = select(a = var_45_to_fp16, b = x_19_cast_fp16, cond = var_576)[name = string("input_53_cast_fp16")];
+            bool new_x_3_interleave_0 = const()[name = string("new_x_3_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_3_cast_fp16 = concat(axis = var_60, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = string("new_x_3_cast_fp16")];
+            tensor<int32, [3]> var_589_begin_0 = const()[name = string("op_589_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_589_end_0 = const()[name = string("op_589_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_589_end_mask_0 = const()[name = string("op_589_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_589_cast_fp16 = slice_by_index(begin = var_589_begin_0, end = var_589_end_0, end_mask = var_589_end_mask_0, x = new_x_3_cast_fp16)[name = string("op_589_cast_fp16")];
+            string x_21_pad_type_0 = const()[name = string("x_21_pad_type_0"), val = string("valid")];
+            int32 x_21_groups_0 = const()[name = string("x_21_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_21_strides_0 = const()[name = string("x_21_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_21_pad_0 = const()[name = string("x_21_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_21_dilations_0 = const()[name = string("x_21_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19459840))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19469120))))[name = string("encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_3_cast_fp16)[name = string("x_21_cast_fp16")];
+            tensor<int32, [3]> input_55_perm_0 = const()[name = string("input_55_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_23_axes_0 = const()[name = string("x_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19471232)))];
+            tensor<fp16, [1024]> encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19473344)))];
+            tensor<fp16, [1, 14, 1024]> input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = string("transpose_356")];
+            tensor<fp16, [1, 14, 1024]> x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = string("x_23_cast_fp16")];
+            tensor<int32, [3]> input_57_perm_0 = const()[name = string("input_57_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = string("transpose_355")];
+            tensor<fp16, [1, 1024, 14]> input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = string("input_59_cast_fp16")];
+            string x_25_pad_type_0 = const()[name = string("x_25_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_25_strides_0 = const()[name = string("x_25_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_25_pad_0 = const()[name = string("x_25_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_25_dilations_0 = const()[name = string("x_25_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_25_groups_0 = const()[name = string("x_25_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19475456))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20524096))))[name = string("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("x_25_cast_fp16")];
+            tensor<int32, [3]> input_61_perm_0 = const()[name = string("input_61_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = string("transpose_354")];
+            tensor<fp16, [1, 14, 1024]> input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = string("input_63_cast_fp16")];
+            tensor<int32, [1]> input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20526208)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20528320)))];
+            tensor<fp16, [1, 14, 1024]> input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20530432))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24724800))))[name = string("encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_0_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24733056)))];
+            tensor<fp16, [1, 14, 4096]> linear_8_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_8_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = string("input_69_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24741312))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28935680))))[name = string("encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28937792)))];
+            tensor<fp16, [1, 14, 1024]> linear_9_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_9_cast_fp16")];
+            fp16 var_632_to_fp16 = const()[name = string("op_632_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_633_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_632_to_fp16)[name = string("op_633_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_633_cast_fp16)[name = string("input_75_cast_fp16")];
+            tensor<int32, [1]> input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28939904)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28942016)))];
+            tensor<fp16, [1, 14, 1024]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")];
+            tensor<int32, [4]> cache_5_begin_0 = const()[name = string("cache_5_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_5_end_0 = const()[name = string("cache_5_end_0"), val = tensor<int32, [4]>([2, 1, 42, 1024])];
+            tensor<bool, [4]> cache_5_end_mask_0 = const()[name = string("cache_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_5_squeeze_mask_0 = const()[name = string("cache_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_5_cast_fp16 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_5_cast_fp16")];
+            tensor<int32, [4]> cache_7_begin_0 = const()[name = string("cache_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_7_end_0 = const()[name = string("cache_7_end_0"), val = tensor<int32, [4]>([2, 1, 1024, 8])];
+            tensor<bool, [4]> cache_7_end_mask_0 = const()[name = string("cache_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_7_squeeze_mask_0 = const()[name = string("cache_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_7_cast_fp16 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_7_cast_fp16")];
+            tensor<int32, [1]> input_79_axes_0 = const()[name = string("input_79_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28944128)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28946240)))];
+            tensor<fp16, [1, 14, 1024]> input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = string("input_79_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28948352))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33142720))))[name = string("encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_1_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33150976)))];
+            tensor<fp16, [1, 14, 4096]> linear_10_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_10_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = string("input_83_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33159232))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37353600))))[name = string("encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37355712)))];
+            tensor<fp16, [1, 14, 1024]> linear_11_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_11_cast_fp16")];
+            fp16 var_669_to_fp16 = const()[name = string("op_669_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_670_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_669_to_fp16)[name = string("op_670_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_670_cast_fp16)[name = string("input_89_cast_fp16")];
+            tensor<int32, [1]> key_3_axes_0 = const()[name = string("key_3_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37357824)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37359936)))];
+            tensor<fp16, [1, 14, 1024]> key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = string("key_3_cast_fp16")];
+            bool input_91_interleave_0 = const()[name = string("input_91_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_91_cast_fp16 = concat(axis = var_69, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = string("input_91_cast_fp16")];
+            tensor<int32, [3]> var_692_begin_0 = const()[name = string("op_692_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_692_end_0 = const()[name = string("op_692_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_692_end_mask_0 = const()[name = string("op_692_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_692_cast_fp16 = slice_by_index(begin = var_692_begin_0, end = var_692_end_0, end_mask = var_692_end_mask_0, x = cache_5_cast_fp16)[name = string("op_692_cast_fp16")];
+            bool var_698_interleave_0 = const()[name = string("op_698_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_698_cast_fp16 = concat(axis = var_69, interleave = var_698_interleave_0, values = (var_692_cast_fp16, key_3_cast_fp16))[name = string("op_698_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37362048))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38410688))))[name = string("encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38412800)))];
+            tensor<fp16, [1, 14, 1024]> linear_12_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized, x = key_3_cast_fp16)[name = string("linear_12_cast_fp16")];
+            tensor<int32, [4]> var_703 = const()[name = string("op_703"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_7_cast_fp16 = reshape(shape = var_703, x = linear_12_cast_fp16)[name = string("q_7_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38414912))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39463552))))[name = string("encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39465664)))];
+            tensor<fp16, [1, 56, 1024]> linear_13_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_13_cast_fp16")];
+            tensor<int32, [4]> var_708 = const()[name = string("op_708"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_5_cast_fp16 = reshape(shape = var_708, x = linear_13_cast_fp16)[name = string("k_5_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39467776))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40516416))))[name = string("encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40518528)))];
+            tensor<fp16, [1, 56, 1024]> linear_14_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_14_cast_fp16")];
+            tensor<int32, [4]> var_713 = const()[name = string("op_713"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_3_cast_fp16 = reshape(shape = var_713, x = linear_14_cast_fp16)[name = string("v_3_cast_fp16")];
+            tensor<int32, [4]> value_11_perm_0 = const()[name = string("value_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40520640)))];
+            tensor<fp16, [1, 14, 8, 128]> var_726_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = string("op_726_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40522752)))];
+            tensor<fp16, [1, 14, 8, 128]> var_728_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = string("op_728_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_3_perm_0 = const()[name = string("q_with_bias_v_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_33_transpose_x_0 = const()[name = string("x_33_transpose_x_0"), val = bool(false)];
+            bool x_33_transpose_y_0 = const()[name = string("x_33_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_730_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40524864))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40638592))))[name = string("op_730_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_728_cast_fp16)[name = string("transpose_353")];
+            tensor<fp16, [1, 8, 14, 111]> x_33_cast_fp16 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = op_730_to_fp16_quantized)[name = string("x_33_cast_fp16")];
+            tensor<int32, [8]> x_35_pad_0 = const()[name = string("x_35_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_35_mode_0 = const()[name = string("x_35_mode_0"), val = string("constant")];
+            fp16 const_92_to_fp16 = const()[name = string("const_92_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_35_cast_fp16 = pad(constant_val = const_92_to_fp16, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33_cast_fp16)[name = string("x_35_cast_fp16")];
+            tensor<int32, [4]> var_738 = const()[name = string("op_738"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_37_cast_fp16 = reshape(shape = var_738, x = x_35_cast_fp16)[name = string("x_37_cast_fp16")];
+            tensor<int32, [4]> var_742_begin_0 = const()[name = string("op_742_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_742_end_0 = const()[name = string("op_742_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_742_end_mask_0 = const()[name = string("op_742_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_742_cast_fp16 = slice_by_index(begin = var_742_begin_0, end = var_742_end_0, end_mask = var_742_end_mask_0, x = x_37_cast_fp16)[name = string("op_742_cast_fp16")];
+            tensor<int32, [4]> var_743 = const()[name = string("op_743"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_5_cast_fp16 = reshape(shape = var_743, x = var_742_cast_fp16)[name = string("matrix_bd_5_cast_fp16")];
+            bool matrix_ac_3_transpose_x_0 = const()[name = string("matrix_ac_3_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_3_transpose_y_0 = const()[name = string("matrix_ac_3_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_99 = transpose(perm = transpose_99_perm_0, x = k_5_cast_fp16)[name = string("transpose_351")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_98 = transpose(perm = transpose_98_perm_0, x = var_726_cast_fp16)[name = string("transpose_352")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_98, y = transpose_99)[name = string("matrix_ac_3_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_7_begin_0 = const()[name = string("matrix_bd_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_7_end_0 = const()[name = string("matrix_bd_7_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_7_end_mask_0 = const()[name = string("matrix_bd_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5_cast_fp16)[name = string("matrix_bd_7_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_752_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = string("op_752_cast_fp16")];
+            fp16 _inversed_scores_5_y_0_to_fp16 = const()[name = string("_inversed_scores_5_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_5_cast_fp16 = mul(x = var_752_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = string("_inversed_scores_5_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_7_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_11)[name = string("scores_7_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_758_cast_fp16 = softmax(axis = var_60, x = scores_7_cast_fp16)[name = string("op_758_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_93_cast_fp16 = select(a = var_45_to_fp16, b = var_758_cast_fp16, cond = mask_11)[name = string("input_93_cast_fp16")];
+            bool x_39_transpose_x_0 = const()[name = string("x_39_transpose_x_0"), val = bool(false)];
+            bool x_39_transpose_y_0 = const()[name = string("x_39_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_3_cast_fp16)[name = string("transpose_350")];
+            tensor<fp16, [1, 8, 14, 128]> x_39_cast_fp16 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93_cast_fp16, y = value_11_cast_fp16)[name = string("x_39_cast_fp16")];
+            tensor<int32, [4]> var_762_perm_0 = const()[name = string("op_762_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_763 = const()[name = string("op_763"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_762_cast_fp16 = transpose(perm = var_762_perm_0, x = x_39_cast_fp16)[name = string("transpose_349")];
+            tensor<fp16, [1, 14, 1024]> input_95_cast_fp16 = reshape(shape = var_763, x = var_762_cast_fp16)[name = string("input_95_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40638912))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41687552))))[name = string("encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41689664)))];
+            tensor<fp16, [1, 14, 1024]> linear_16_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_16_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = string("input_99_cast_fp16")];
+            tensor<int32, [1]> x_43_axes_0 = const()[name = string("x_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41691776)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41693888)))];
+            tensor<fp16, [1, 14, 1024]> x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = string("x_43_cast_fp16")];
+            tensor<int32, [3]> input_101_perm_0 = const()[name = string("input_101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_103_pad_type_0 = const()[name = string("input_103_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_103_strides_0 = const()[name = string("input_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_103_pad_0 = const()[name = string("input_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_103_dilations_0 = const()[name = string("input_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_103_groups_0 = const()[name = string("input_103_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41696000))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43793216))))[name = string("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = string("transpose_348")];
+            tensor<fp16, [1, 2048, 14]> input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")];
+            int32 x_45_split_num_splits_0 = const()[name = string("x_45_split_num_splits_0"), val = int32(2)];
+            int32 x_45_split_axis_0 = const()[name = string("x_45_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_45_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_45_split_cast_fp16_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103_cast_fp16)[name = string("x_45_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = string("x_45_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = string("x_45_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_105_cast_fp16 = select(a = var_45_to_fp16, b = x_45_cast_fp16, cond = var_576)[name = string("input_105_cast_fp16")];
+            bool new_x_7_interleave_0 = const()[name = string("new_x_7_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_7_cast_fp16 = concat(axis = var_60, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = string("new_x_7_cast_fp16")];
+            tensor<int32, [3]> var_802_begin_0 = const()[name = string("op_802_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_802_end_0 = const()[name = string("op_802_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_802_end_mask_0 = const()[name = string("op_802_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_802_cast_fp16 = slice_by_index(begin = var_802_begin_0, end = var_802_end_0, end_mask = var_802_end_mask_0, x = new_x_7_cast_fp16)[name = string("op_802_cast_fp16")];
+            string x_47_pad_type_0 = const()[name = string("x_47_pad_type_0"), val = string("valid")];
+            int32 x_47_groups_0 = const()[name = string("x_47_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_47_strides_0 = const()[name = string("x_47_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_47_pad_0 = const()[name = string("x_47_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_47_dilations_0 = const()[name = string("x_47_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43797376))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43806656))))[name = string("encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_7_cast_fp16)[name = string("x_47_cast_fp16")];
+            tensor<int32, [3]> input_107_perm_0 = const()[name = string("input_107_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_49_axes_0 = const()[name = string("x_49_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43808768)))];
+            tensor<fp16, [1024]> encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43810880)))];
+            tensor<fp16, [1, 14, 1024]> input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = string("transpose_347")];
+            tensor<fp16, [1, 14, 1024]> x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = string("x_49_cast_fp16")];
+            tensor<int32, [3]> input_109_perm_0 = const()[name = string("input_109_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = string("transpose_346")];
+            tensor<fp16, [1, 1024, 14]> input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = string("input_111_cast_fp16")];
+            string x_51_pad_type_0 = const()[name = string("x_51_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_51_strides_0 = const()[name = string("x_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_51_pad_0 = const()[name = string("x_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_51_dilations_0 = const()[name = string("x_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_51_groups_0 = const()[name = string("x_51_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43812992))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44861632))))[name = string("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("x_51_cast_fp16")];
+            tensor<int32, [3]> input_113_perm_0 = const()[name = string("input_113_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = string("transpose_345")];
+            tensor<fp16, [1, 14, 1024]> input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = string("input_115_cast_fp16")];
+            tensor<int32, [1]> input_117_axes_0 = const()[name = string("input_117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44863744)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44865856)))];
+            tensor<fp16, [1, 14, 1024]> input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = string("input_117_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44867968))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49062336))))[name = string("encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_1_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49070592)))];
+            tensor<fp16, [1, 14, 4096]> linear_17_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_17_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = string("input_121_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49078848))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53273216))))[name = string("encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53275328)))];
+            tensor<fp16, [1, 14, 1024]> linear_18_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_18_cast_fp16")];
+            fp16 var_845_to_fp16 = const()[name = string("op_845_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_846_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_845_to_fp16)[name = string("op_846_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_846_cast_fp16)[name = string("input_127_cast_fp16")];
+            tensor<int32, [1]> input_129_axes_0 = const()[name = string("input_129_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53277440)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53279552)))];
+            tensor<fp16, [1, 14, 1024]> input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = string("input_129_cast_fp16")];
+            tensor<int32, [4]> cache_9_begin_0 = const()[name = string("cache_9_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_9_end_0 = const()[name = string("cache_9_end_0"), val = tensor<int32, [4]>([3, 1, 42, 1024])];
+            tensor<bool, [4]> cache_9_end_mask_0 = const()[name = string("cache_9_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_9_squeeze_mask_0 = const()[name = string("cache_9_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_9_cast_fp16 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_9_cast_fp16")];
+            tensor<int32, [4]> cache_11_begin_0 = const()[name = string("cache_11_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_11_end_0 = const()[name = string("cache_11_end_0"), val = tensor<int32, [4]>([3, 1, 1024, 8])];
+            tensor<bool, [4]> cache_11_end_mask_0 = const()[name = string("cache_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_11_squeeze_mask_0 = const()[name = string("cache_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_11_cast_fp16 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_11_cast_fp16")];
+            tensor<int32, [1]> input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53281664)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53283776)))];
+            tensor<fp16, [1, 14, 1024]> input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53285888))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57480256))))[name = string("encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_2_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57488512)))];
+            tensor<fp16, [1, 14, 4096]> linear_19_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_19_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = string("input_135_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57496768))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61691136))))[name = string("encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61693248)))];
+            tensor<fp16, [1, 14, 1024]> linear_20_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_20_cast_fp16")];
+            fp16 var_882_to_fp16 = const()[name = string("op_882_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_883_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_882_to_fp16)[name = string("op_883_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_883_cast_fp16)[name = string("input_141_cast_fp16")];
+            tensor<int32, [1]> key_5_axes_0 = const()[name = string("key_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61695360)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61697472)))];
+            tensor<fp16, [1, 14, 1024]> key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = string("key_5_cast_fp16")];
+            bool input_143_interleave_0 = const()[name = string("input_143_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_143_cast_fp16 = concat(axis = var_69, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = string("input_143_cast_fp16")];
+            tensor<int32, [3]> var_905_begin_0 = const()[name = string("op_905_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_905_end_0 = const()[name = string("op_905_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_905_end_mask_0 = const()[name = string("op_905_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_905_cast_fp16 = slice_by_index(begin = var_905_begin_0, end = var_905_end_0, end_mask = var_905_end_mask_0, x = cache_9_cast_fp16)[name = string("op_905_cast_fp16")];
+            bool var_911_interleave_0 = const()[name = string("op_911_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_911_cast_fp16 = concat(axis = var_69, interleave = var_911_interleave_0, values = (var_905_cast_fp16, key_5_cast_fp16))[name = string("op_911_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61699584))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62748224))))[name = string("encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62750336)))];
+            tensor<fp16, [1, 14, 1024]> linear_21_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized, x = key_5_cast_fp16)[name = string("linear_21_cast_fp16")];
+            tensor<int32, [4]> var_916 = const()[name = string("op_916"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_13_cast_fp16 = reshape(shape = var_916, x = linear_21_cast_fp16)[name = string("q_13_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62752448))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63801088))))[name = string("encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63803200)))];
+            tensor<fp16, [1, 56, 1024]> linear_22_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_22_cast_fp16")];
+            tensor<int32, [4]> var_921 = const()[name = string("op_921"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_9_cast_fp16 = reshape(shape = var_921, x = linear_22_cast_fp16)[name = string("k_9_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63805312))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64853952))))[name = string("encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64856064)))];
+            tensor<fp16, [1, 56, 1024]> linear_23_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_23_cast_fp16")];
+            tensor<int32, [4]> var_926 = const()[name = string("op_926"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_5_cast_fp16 = reshape(shape = var_926, x = linear_23_cast_fp16)[name = string("v_5_cast_fp16")];
+            tensor<int32, [4]> value_13_perm_0 = const()[name = string("value_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64858176)))];
+            tensor<fp16, [1, 14, 8, 128]> var_939_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = string("op_939_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64860288)))];
+            tensor<fp16, [1, 14, 8, 128]> var_941_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = string("op_941_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_5_perm_0 = const()[name = string("q_with_bias_v_5_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_59_transpose_x_0 = const()[name = string("x_59_transpose_x_0"), val = bool(false)];
+            bool x_59_transpose_y_0 = const()[name = string("x_59_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_943_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64862400))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64976128))))[name = string("op_943_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_941_cast_fp16)[name = string("transpose_344")];
+            tensor<fp16, [1, 8, 14, 111]> x_59_cast_fp16 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = op_943_to_fp16_quantized)[name = string("x_59_cast_fp16")];
+            tensor<int32, [8]> x_61_pad_0 = const()[name = string("x_61_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_61_mode_0 = const()[name = string("x_61_mode_0"), val = string("constant")];
+            fp16 const_105_to_fp16 = const()[name = string("const_105_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_61_cast_fp16 = pad(constant_val = const_105_to_fp16, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59_cast_fp16)[name = string("x_61_cast_fp16")];
+            tensor<int32, [4]> var_951 = const()[name = string("op_951"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_63_cast_fp16 = reshape(shape = var_951, x = x_61_cast_fp16)[name = string("x_63_cast_fp16")];
+            tensor<int32, [4]> var_955_begin_0 = const()[name = string("op_955_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_955_end_0 = const()[name = string("op_955_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_955_end_mask_0 = const()[name = string("op_955_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_955_cast_fp16 = slice_by_index(begin = var_955_begin_0, end = var_955_end_0, end_mask = var_955_end_mask_0, x = x_63_cast_fp16)[name = string("op_955_cast_fp16")];
+            tensor<int32, [4]> var_956 = const()[name = string("op_956"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_9_cast_fp16 = reshape(shape = var_956, x = var_955_cast_fp16)[name = string("matrix_bd_9_cast_fp16")];
+            bool matrix_ac_5_transpose_x_0 = const()[name = string("matrix_ac_5_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_5_transpose_y_0 = const()[name = string("matrix_ac_5_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_101 = transpose(perm = transpose_101_perm_0, x = k_9_cast_fp16)[name = string("transpose_342")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_100 = transpose(perm = transpose_100_perm_0, x = var_939_cast_fp16)[name = string("transpose_343")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_100, y = transpose_101)[name = string("matrix_ac_5_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_11_begin_0 = const()[name = string("matrix_bd_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_11_end_0 = const()[name = string("matrix_bd_11_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_11_end_mask_0 = const()[name = string("matrix_bd_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9_cast_fp16)[name = string("matrix_bd_11_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_965_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = string("op_965_cast_fp16")];
+            fp16 _inversed_scores_9_y_0_to_fp16 = const()[name = string("_inversed_scores_9_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_9_cast_fp16 = mul(x = var_965_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = string("_inversed_scores_9_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_11_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_11)[name = string("scores_11_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_971_cast_fp16 = softmax(axis = var_60, x = scores_11_cast_fp16)[name = string("op_971_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_145_cast_fp16 = select(a = var_45_to_fp16, b = var_971_cast_fp16, cond = mask_11)[name = string("input_145_cast_fp16")];
+            bool x_65_transpose_x_0 = const()[name = string("x_65_transpose_x_0"), val = bool(false)];
+            bool x_65_transpose_y_0 = const()[name = string("x_65_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_5_cast_fp16)[name = string("transpose_341")];
+            tensor<fp16, [1, 8, 14, 128]> x_65_cast_fp16 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145_cast_fp16, y = value_13_cast_fp16)[name = string("x_65_cast_fp16")];
+            tensor<int32, [4]> var_975_perm_0 = const()[name = string("op_975_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_976 = const()[name = string("op_976"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_975_cast_fp16 = transpose(perm = var_975_perm_0, x = x_65_cast_fp16)[name = string("transpose_340")];
+            tensor<fp16, [1, 14, 1024]> input_147_cast_fp16 = reshape(shape = var_976, x = var_975_cast_fp16)[name = string("input_147_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64976448))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65762944))))[name = string("encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65763136)))];
+            tensor<fp16, [1, 14, 1024]> linear_25_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = string("linear_25_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = string("input_151_cast_fp16")];
+            tensor<int32, [1]> x_69_axes_0 = const()[name = string("x_69_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65765248)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65767360)))];
+            tensor<fp16, [1, 14, 1024]> x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = string("x_69_cast_fp16")];
+            tensor<int32, [3]> input_153_perm_0 = const()[name = string("input_153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_155_pad_type_0 = const()[name = string("input_155_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_155_strides_0 = const()[name = string("input_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_155_pad_0 = const()[name = string("input_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_155_dilations_0 = const()[name = string("input_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_155_groups_0 = const()[name = string("input_155_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65769472))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67866688))))[name = string("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = string("transpose_339")];
+            tensor<fp16, [1, 2048, 14]> input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")];
+            int32 x_71_split_num_splits_0 = const()[name = string("x_71_split_num_splits_0"), val = int32(2)];
+            int32 x_71_split_axis_0 = const()[name = string("x_71_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_71_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_71_split_cast_fp16_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155_cast_fp16)[name = string("x_71_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = string("x_71_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = string("x_71_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_157_cast_fp16 = select(a = var_45_to_fp16, b = x_71_cast_fp16, cond = var_576)[name = string("input_157_cast_fp16")];
+            bool new_x_11_interleave_0 = const()[name = string("new_x_11_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_11_cast_fp16 = concat(axis = var_60, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = string("new_x_11_cast_fp16")];
+            tensor<int32, [3]> var_1015_begin_0 = const()[name = string("op_1015_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_1015_end_0 = const()[name = string("op_1015_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_1015_end_mask_0 = const()[name = string("op_1015_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1015_cast_fp16 = slice_by_index(begin = var_1015_begin_0, end = var_1015_end_0, end_mask = var_1015_end_mask_0, x = new_x_11_cast_fp16)[name = string("op_1015_cast_fp16")];
+            string x_73_pad_type_0 = const()[name = string("x_73_pad_type_0"), val = string("valid")];
+            int32 x_73_groups_0 = const()[name = string("x_73_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_73_strides_0 = const()[name = string("x_73_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_73_pad_0 = const()[name = string("x_73_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_73_dilations_0 = const()[name = string("x_73_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67870848))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67880128))))[name = string("encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_73_cast_fp16 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_11_cast_fp16)[name = string("x_73_cast_fp16")];
+            tensor<int32, [3]> input_159_perm_0 = const()[name = string("input_159_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_75_axes_0 = const()[name = string("x_75_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67882240)))];
+            tensor<fp16, [1024]> encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67884352)))];
+            tensor<fp16, [1, 14, 1024]> input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = string("transpose_338")];
+            tensor<fp16, [1, 14, 1024]> x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = string("x_75_cast_fp16")];
+            tensor<int32, [3]> input_161_perm_0 = const()[name = string("input_161_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = string("transpose_337")];
+            tensor<fp16, [1, 1024, 14]> input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = string("input_163_cast_fp16")];
+            string x_77_pad_type_0 = const()[name = string("x_77_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_77_strides_0 = const()[name = string("x_77_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_77_pad_0 = const()[name = string("x_77_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_77_dilations_0 = const()[name = string("x_77_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_77_groups_0 = const()[name = string("x_77_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67886464))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68935104))))[name = string("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("x_77_cast_fp16")];
+            tensor<int32, [3]> input_165_perm_0 = const()[name = string("input_165_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = string("transpose_336")];
+            tensor<fp16, [1, 14, 1024]> input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = string("input_167_cast_fp16")];
+            tensor<int32, [1]> input_169_axes_0 = const()[name = string("input_169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68937216)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68939328)))];
+            tensor<fp16, [1, 14, 1024]> input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = string("input_169_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68941440))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72087232))))[name = string("encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_2_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72087424)))];
+            tensor<fp16, [1, 14, 4096]> linear_26_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = string("linear_26_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = string("input_173_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72095680))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75241472))))[name = string("encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_2_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75241664)))];
+            tensor<fp16, [1, 14, 1024]> linear_27_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized, x = input_173_cast_fp16)[name = string("linear_27_cast_fp16")];
+            fp16 var_1058_to_fp16 = const()[name = string("op_1058_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1059_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1058_to_fp16)[name = string("op_1059_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_1059_cast_fp16)[name = string("input_179_cast_fp16")];
+            tensor<int32, [1]> input_181_axes_0 = const()[name = string("input_181_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75243776)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75245888)))];
+            tensor<fp16, [1, 14, 1024]> input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = string("input_181_cast_fp16")];
+            tensor<int32, [4]> cache_13_begin_0 = const()[name = string("cache_13_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_13_end_0 = const()[name = string("cache_13_end_0"), val = tensor<int32, [4]>([4, 1, 42, 1024])];
+            tensor<bool, [4]> cache_13_end_mask_0 = const()[name = string("cache_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_13_squeeze_mask_0 = const()[name = string("cache_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_13_cast_fp16 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_13_cast_fp16")];
+            tensor<int32, [4]> cache_15_begin_0 = const()[name = string("cache_15_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_15_end_0 = const()[name = string("cache_15_end_0"), val = tensor<int32, [4]>([4, 1, 1024, 8])];
+            tensor<bool, [4]> cache_15_end_mask_0 = const()[name = string("cache_15_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_15_squeeze_mask_0 = const()[name = string("cache_15_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_15_cast_fp16 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_15_cast_fp16")];
+            tensor<int32, [1]> input_183_axes_0 = const()[name = string("input_183_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75248000)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75250112)))];
+            tensor<fp16, [1, 14, 1024]> input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = string("input_183_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75252224))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78398016))))[name = string("encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_3_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78398208)))];
+            tensor<fp16, [1, 14, 4096]> linear_28_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = string("linear_28_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = string("input_187_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78406464))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81552256))))[name = string("encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81552448)))];
+            tensor<fp16, [1, 14, 1024]> linear_29_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = string("linear_29_cast_fp16")];
+            fp16 var_1095_to_fp16 = const()[name = string("op_1095_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1096_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_1095_to_fp16)[name = string("op_1096_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_1096_cast_fp16)[name = string("input_193_cast_fp16")];
+            tensor<int32, [1]> key_7_axes_0 = const()[name = string("key_7_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81554560)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81556672)))];
+            tensor<fp16, [1, 14, 1024]> key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = string("key_7_cast_fp16")];
+            bool input_195_interleave_0 = const()[name = string("input_195_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_195_cast_fp16 = concat(axis = var_69, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = string("input_195_cast_fp16")];
+            tensor<int32, [3]> var_1118_begin_0 = const()[name = string("op_1118_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_1118_end_0 = const()[name = string("op_1118_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1118_end_mask_0 = const()[name = string("op_1118_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_1118_cast_fp16 = slice_by_index(begin = var_1118_begin_0, end = var_1118_end_0, end_mask = var_1118_end_mask_0, x = cache_13_cast_fp16)[name = string("op_1118_cast_fp16")];
+            bool var_1124_interleave_0 = const()[name = string("op_1124_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1124_cast_fp16 = concat(axis = var_69, interleave = var_1124_interleave_0, values = (var_1118_cast_fp16, key_7_cast_fp16))[name = string("op_1124_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81558784))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82345280))))[name = string("encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82345472)))];
+            tensor<fp16, [1, 14, 1024]> linear_30_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized, x = key_7_cast_fp16)[name = string("linear_30_cast_fp16")];
+            tensor<int32, [4]> var_1129 = const()[name = string("op_1129"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_19_cast_fp16 = reshape(shape = var_1129, x = linear_30_cast_fp16)[name = string("q_19_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82347584))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83134080))))[name = string("encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83134272)))];
+            tensor<fp16, [1, 56, 1024]> linear_31_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_31_cast_fp16")];
+            tensor<int32, [4]> var_1134 = const()[name = string("op_1134"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_13_cast_fp16 = reshape(shape = var_1134, x = linear_31_cast_fp16)[name = string("k_13_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83136384))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83922880))))[name = string("encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83923072)))];
+            tensor<fp16, [1, 56, 1024]> linear_32_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_32_cast_fp16")];
+            tensor<int32, [4]> var_1139 = const()[name = string("op_1139"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_7_cast_fp16 = reshape(shape = var_1139, x = linear_32_cast_fp16)[name = string("v_7_cast_fp16")];
+            tensor<int32, [4]> value_15_perm_0 = const()[name = string("value_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83925184)))];
+            tensor<fp16, [1, 14, 8, 128]> var_1152_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = string("op_1152_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83927296)))];
+            tensor<fp16, [1, 14, 8, 128]> var_1154_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = string("op_1154_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_7_perm_0 = const()[name = string("q_with_bias_v_7_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_85_transpose_x_0 = const()[name = string("x_85_transpose_x_0"), val = bool(false)];
+            bool x_85_transpose_y_0 = const()[name = string("x_85_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_1156_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83929408))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84043136))))[name = string("op_1156_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_1154_cast_fp16)[name = string("transpose_335")];
+            tensor<fp16, [1, 8, 14, 111]> x_85_cast_fp16 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = op_1156_to_fp16_quantized)[name = string("x_85_cast_fp16")];
+            tensor<int32, [8]> x_87_pad_0 = const()[name = string("x_87_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("constant")];
+            fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_87_cast_fp16 = pad(constant_val = const_118_to_fp16, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85_cast_fp16)[name = string("x_87_cast_fp16")];
+            tensor<int32, [4]> var_1164 = const()[name = string("op_1164"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_89_cast_fp16 = reshape(shape = var_1164, x = x_87_cast_fp16)[name = string("x_89_cast_fp16")];
+            tensor<int32, [4]> var_1168_begin_0 = const()[name = string("op_1168_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1168_end_0 = const()[name = string("op_1168_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_1168_end_mask_0 = const()[name = string("op_1168_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_1168_cast_fp16 = slice_by_index(begin = var_1168_begin_0, end = var_1168_end_0, end_mask = var_1168_end_mask_0, x = x_89_cast_fp16)[name = string("op_1168_cast_fp16")];
+            tensor<int32, [4]> var_1169 = const()[name = string("op_1169"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_13_cast_fp16 = reshape(shape = var_1169, x = var_1168_cast_fp16)[name = string("matrix_bd_13_cast_fp16")];
+            bool matrix_ac_7_transpose_x_0 = const()[name = string("matrix_ac_7_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_7_transpose_y_0 = const()[name = string("matrix_ac_7_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_103 = transpose(perm = transpose_103_perm_0, x = k_13_cast_fp16)[name = string("transpose_333")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_102 = transpose(perm = transpose_102_perm_0, x = var_1152_cast_fp16)[name = string("transpose_334")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_102, y = transpose_103)[name = string("matrix_ac_7_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_15_begin_0 = const()[name = string("matrix_bd_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_15_end_0 = const()[name = string("matrix_bd_15_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_15_end_mask_0 = const()[name = string("matrix_bd_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13_cast_fp16)[name = string("matrix_bd_15_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_1178_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = string("op_1178_cast_fp16")];
+            fp16 _inversed_scores_13_y_0_to_fp16 = const()[name = string("_inversed_scores_13_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_13_cast_fp16 = mul(x = var_1178_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = string("_inversed_scores_13_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_15_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_11)[name = string("scores_15_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_1184_cast_fp16 = softmax(axis = var_60, x = scores_15_cast_fp16)[name = string("op_1184_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_197_cast_fp16 = select(a = var_45_to_fp16, b = var_1184_cast_fp16, cond = mask_11)[name = string("input_197_cast_fp16")];
+            bool x_91_transpose_x_0 = const()[name = string("x_91_transpose_x_0"), val = bool(false)];
+            bool x_91_transpose_y_0 = const()[name = string("x_91_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_7_cast_fp16)[name = string("transpose_332")];
+            tensor<fp16, [1, 8, 14, 128]> x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197_cast_fp16, y = value_15_cast_fp16)[name = string("x_91_cast_fp16")];
+            tensor<int32, [4]> var_1188_perm_0 = const()[name = string("op_1188_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1189 = const()[name = string("op_1189"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_1188_cast_fp16 = transpose(perm = var_1188_perm_0, x = x_91_cast_fp16)[name = string("transpose_331")];
+            tensor<fp16, [1, 14, 1024]> input_199_cast_fp16 = reshape(shape = var_1189, x = var_1188_cast_fp16)[name = string("input_199_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84043456))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84829952))))[name = string("encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84830144)))];
+            tensor<fp16, [1, 14, 1024]> linear_34_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = string("linear_34_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = string("input_203_cast_fp16")];
+            tensor<int32, [1]> x_95_axes_0 = const()[name = string("x_95_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84832256)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84834368)))];
+            tensor<fp16, [1, 14, 1024]> x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = string("x_95_cast_fp16")];
+            tensor<int32, [3]> input_205_perm_0 = const()[name = string("input_205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_207_pad_type_0 = const()[name = string("input_207_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_207_strides_0 = const()[name = string("input_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_207_pad_0 = const()[name = string("input_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_207_dilations_0 = const()[name = string("input_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_207_groups_0 = const()[name = string("input_207_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84836480))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86933696))))[name = string("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = string("transpose_330")];
+            tensor<fp16, [1, 2048, 14]> input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("input_207_cast_fp16")];
+            int32 x_97_split_num_splits_0 = const()[name = string("x_97_split_num_splits_0"), val = int32(2)];
+            int32 x_97_split_axis_0 = const()[name = string("x_97_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_97_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207_cast_fp16)[name = string("x_97_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = string("x_97_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = string("x_97_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_209_cast_fp16 = select(a = var_45_to_fp16, b = x_97_cast_fp16, cond = var_576)[name = string("input_209_cast_fp16")];
+            bool new_x_15_interleave_0 = const()[name = string("new_x_15_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_15_cast_fp16 = concat(axis = var_60, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = string("new_x_15_cast_fp16")];
+            tensor<int32, [3]> var_1228_begin_0 = const()[name = string("op_1228_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_1228_end_0 = const()[name = string("op_1228_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_1228_end_mask_0 = const()[name = string("op_1228_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1228_cast_fp16 = slice_by_index(begin = var_1228_begin_0, end = var_1228_end_0, end_mask = var_1228_end_mask_0, x = new_x_15_cast_fp16)[name = string("op_1228_cast_fp16")];
+            string x_99_pad_type_0 = const()[name = string("x_99_pad_type_0"), val = string("valid")];
+            int32 x_99_groups_0 = const()[name = string("x_99_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_99_strides_0 = const()[name = string("x_99_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_99_pad_0 = const()[name = string("x_99_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_99_dilations_0 = const()[name = string("x_99_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86937856))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86947136))))[name = string("encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_15_cast_fp16)[name = string("x_99_cast_fp16")];
+            tensor<int32, [3]> input_211_perm_0 = const()[name = string("input_211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_101_axes_0 = const()[name = string("x_101_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86949248)))];
+            tensor<fp16, [1024]> encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86951360)))];
+            tensor<fp16, [1, 14, 1024]> input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = string("transpose_329")];
+            tensor<fp16, [1, 14, 1024]> x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = string("x_101_cast_fp16")];
+            tensor<int32, [3]> input_213_perm_0 = const()[name = string("input_213_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = string("transpose_328")];
+            tensor<fp16, [1, 1024, 14]> input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = string("input_215_cast_fp16")];
+            string x_103_pad_type_0 = const()[name = string("x_103_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_103_strides_0 = const()[name = string("x_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_103_pad_0 = const()[name = string("x_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_103_dilations_0 = const()[name = string("x_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_103_groups_0 = const()[name = string("x_103_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86953472))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88002112))))[name = string("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("x_103_cast_fp16")];
+            tensor<int32, [3]> input_217_perm_0 = const()[name = string("input_217_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = string("transpose_327")];
+            tensor<fp16, [1, 14, 1024]> input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = string("input_219_cast_fp16")];
+            tensor<int32, [1]> input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88004224)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88006336)))];
+            tensor<fp16, [1, 14, 1024]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88008448))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91154240))))[name = string("encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_3_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91154432)))];
+            tensor<fp16, [1, 14, 4096]> linear_35_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = string("linear_35_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = string("input_225_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91162688))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94308480))))[name = string("encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94308672)))];
+            tensor<fp16, [1, 14, 1024]> linear_36_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = string("linear_36_cast_fp16")];
+            fp16 var_1271_to_fp16 = const()[name = string("op_1271_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1272_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1271_to_fp16)[name = string("op_1272_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1272_cast_fp16)[name = string("input_231_cast_fp16")];
+            tensor<int32, [1]> input_233_axes_0 = const()[name = string("input_233_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94310784)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94312896)))];
+            tensor<fp16, [1, 14, 1024]> input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = string("input_233_cast_fp16")];
+            tensor<int32, [4]> cache_17_begin_0 = const()[name = string("cache_17_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_17_end_0 = const()[name = string("cache_17_end_0"), val = tensor<int32, [4]>([5, 1, 42, 1024])];
+            tensor<bool, [4]> cache_17_end_mask_0 = const()[name = string("cache_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_17_squeeze_mask_0 = const()[name = string("cache_17_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_17_cast_fp16 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_17_cast_fp16")];
+            tensor<int32, [4]> cache_19_begin_0 = const()[name = string("cache_19_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_19_end_0 = const()[name = string("cache_19_end_0"), val = tensor<int32, [4]>([5, 1, 1024, 8])];
+            tensor<bool, [4]> cache_19_end_mask_0 = const()[name = string("cache_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_19_squeeze_mask_0 = const()[name = string("cache_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_19_cast_fp16 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_19_cast_fp16")];
+            tensor<int32, [1]> input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94315008)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94317120)))];
+            tensor<fp16, [1, 14, 1024]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94319232))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97465024))))[name = string("encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_4_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97465216)))];
+            tensor<fp16, [1, 14, 4096]> linear_37_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized, x = input_235_cast_fp16)[name = string("linear_37_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = string("input_239_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97473472))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100619264))))[name = string("encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100619456)))];
+            tensor<fp16, [1, 14, 1024]> linear_38_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = string("linear_38_cast_fp16")];
+            fp16 var_1308_to_fp16 = const()[name = string("op_1308_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1309_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1308_to_fp16)[name = string("op_1309_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1309_cast_fp16)[name = string("input_245_cast_fp16")];
+            tensor<int32, [1]> key_9_axes_0 = const()[name = string("key_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100621568)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100623680)))];
+            tensor<fp16, [1, 14, 1024]> key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = string("key_9_cast_fp16")];
+            bool input_247_interleave_0 = const()[name = string("input_247_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_247_cast_fp16 = concat(axis = var_69, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = string("input_247_cast_fp16")];
+            tensor<int32, [3]> var_1331_begin_0 = const()[name = string("op_1331_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_1331_end_0 = const()[name = string("op_1331_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1331_end_mask_0 = const()[name = string("op_1331_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_1331_cast_fp16 = slice_by_index(begin = var_1331_begin_0, end = var_1331_end_0, end_mask = var_1331_end_mask_0, x = cache_17_cast_fp16)[name = string("op_1331_cast_fp16")];
+            bool var_1337_interleave_0 = const()[name = string("op_1337_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1337_cast_fp16 = concat(axis = var_69, interleave = var_1337_interleave_0, values = (var_1331_cast_fp16, key_9_cast_fp16))[name = string("op_1337_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100625792))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101412288))))[name = string("encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101412480)))];
+            tensor<fp16, [1, 14, 1024]> linear_39_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized, x = key_9_cast_fp16)[name = string("linear_39_cast_fp16")];
+            tensor<int32, [4]> var_1342 = const()[name = string("op_1342"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_25_cast_fp16 = reshape(shape = var_1342, x = linear_39_cast_fp16)[name = string("q_25_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101414592))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102201088))))[name = string("encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102201280)))];
+            tensor<fp16, [1, 56, 1024]> linear_40_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_40_cast_fp16")];
+            tensor<int32, [4]> var_1347 = const()[name = string("op_1347"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_17_cast_fp16 = reshape(shape = var_1347, x = linear_40_cast_fp16)[name = string("k_17_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102203392))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102989888))))[name = string("encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102990080)))];
+            tensor<fp16, [1, 56, 1024]> linear_41_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_41_cast_fp16")];
+            tensor<int32, [4]> var_1352 = const()[name = string("op_1352"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_9_cast_fp16 = reshape(shape = var_1352, x = linear_41_cast_fp16)[name = string("v_9_cast_fp16")];
+            tensor<int32, [4]> value_17_perm_0 = const()[name = string("value_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102992192)))];
+            tensor<fp16, [1, 14, 8, 128]> var_1365_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = string("op_1365_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102994304)))];
+            tensor<fp16, [1, 14, 8, 128]> var_1367_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = string("op_1367_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_9_perm_0 = const()[name = string("q_with_bias_v_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_111_transpose_x_0 = const()[name = string("x_111_transpose_x_0"), val = bool(false)];
+            bool x_111_transpose_y_0 = const()[name = string("x_111_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_1369_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102996416))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103110144))))[name = string("op_1369_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1367_cast_fp16)[name = string("transpose_326")];
+            tensor<fp16, [1, 8, 14, 111]> x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = op_1369_to_fp16_quantized)[name = string("x_111_cast_fp16")];
+            tensor<int32, [8]> x_113_pad_0 = const()[name = string("x_113_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_113_mode_0 = const()[name = string("x_113_mode_0"), val = string("constant")];
+            fp16 const_131_to_fp16 = const()[name = string("const_131_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_113_cast_fp16 = pad(constant_val = const_131_to_fp16, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111_cast_fp16)[name = string("x_113_cast_fp16")];
+            tensor<int32, [4]> var_1377 = const()[name = string("op_1377"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_115_cast_fp16 = reshape(shape = var_1377, x = x_113_cast_fp16)[name = string("x_115_cast_fp16")];
+            tensor<int32, [4]> var_1381_begin_0 = const()[name = string("op_1381_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1381_end_0 = const()[name = string("op_1381_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_1381_end_mask_0 = const()[name = string("op_1381_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_1381_cast_fp16 = slice_by_index(begin = var_1381_begin_0, end = var_1381_end_0, end_mask = var_1381_end_mask_0, x = x_115_cast_fp16)[name = string("op_1381_cast_fp16")];
+            tensor<int32, [4]> var_1382 = const()[name = string("op_1382"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_17_cast_fp16 = reshape(shape = var_1382, x = var_1381_cast_fp16)[name = string("matrix_bd_17_cast_fp16")];
+            bool matrix_ac_9_transpose_x_0 = const()[name = string("matrix_ac_9_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_9_transpose_y_0 = const()[name = string("matrix_ac_9_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_105 = transpose(perm = transpose_105_perm_0, x = k_17_cast_fp16)[name = string("transpose_324")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1365_cast_fp16)[name = string("transpose_325")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_104, y = transpose_105)[name = string("matrix_ac_9_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_19_begin_0 = const()[name = string("matrix_bd_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_19_end_0 = const()[name = string("matrix_bd_19_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_19_end_mask_0 = const()[name = string("matrix_bd_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17_cast_fp16)[name = string("matrix_bd_19_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_1391_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = string("op_1391_cast_fp16")];
+            fp16 _inversed_scores_17_y_0_to_fp16 = const()[name = string("_inversed_scores_17_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_17_cast_fp16 = mul(x = var_1391_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = string("_inversed_scores_17_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_19_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_11)[name = string("scores_19_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_1397_cast_fp16 = softmax(axis = var_60, x = scores_19_cast_fp16)[name = string("op_1397_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_249_cast_fp16 = select(a = var_45_to_fp16, b = var_1397_cast_fp16, cond = mask_11)[name = string("input_249_cast_fp16")];
+            bool x_117_transpose_x_0 = const()[name = string("x_117_transpose_x_0"), val = bool(false)];
+            bool x_117_transpose_y_0 = const()[name = string("x_117_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_9_cast_fp16)[name = string("transpose_323")];
+            tensor<fp16, [1, 8, 14, 128]> x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249_cast_fp16, y = value_17_cast_fp16)[name = string("x_117_cast_fp16")];
+            tensor<int32, [4]> var_1401_perm_0 = const()[name = string("op_1401_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1402 = const()[name = string("op_1402"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_1401_cast_fp16 = transpose(perm = var_1401_perm_0, x = x_117_cast_fp16)[name = string("transpose_322")];
+            tensor<fp16, [1, 14, 1024]> input_251_cast_fp16 = reshape(shape = var_1402, x = var_1401_cast_fp16)[name = string("input_251_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103110464))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103896960))))[name = string("encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103897152)))];
+            tensor<fp16, [1, 14, 1024]> linear_43_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = string("linear_43_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = string("input_255_cast_fp16")];
+            tensor<int32, [1]> x_121_axes_0 = const()[name = string("x_121_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103899264)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103901376)))];
+            tensor<fp16, [1, 14, 1024]> x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = string("x_121_cast_fp16")];
+            tensor<int32, [3]> input_257_perm_0 = const()[name = string("input_257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_259_pad_type_0 = const()[name = string("input_259_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_259_strides_0 = const()[name = string("input_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_259_pad_0 = const()[name = string("input_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_259_dilations_0 = const()[name = string("input_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_259_groups_0 = const()[name = string("input_259_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103903488))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106000704))))[name = string("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = string("transpose_321")];
+            tensor<fp16, [1, 2048, 14]> input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("input_259_cast_fp16")];
+            int32 x_123_split_num_splits_0 = const()[name = string("x_123_split_num_splits_0"), val = int32(2)];
+            int32 x_123_split_axis_0 = const()[name = string("x_123_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_123_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_123_split_cast_fp16_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259_cast_fp16)[name = string("x_123_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = string("x_123_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = string("x_123_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_261_cast_fp16 = select(a = var_45_to_fp16, b = x_123_cast_fp16, cond = var_576)[name = string("input_261_cast_fp16")];
+            bool new_x_19_interleave_0 = const()[name = string("new_x_19_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_19_cast_fp16 = concat(axis = var_60, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = string("new_x_19_cast_fp16")];
+            tensor<int32, [3]> var_1441_begin_0 = const()[name = string("op_1441_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_1441_end_0 = const()[name = string("op_1441_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_1441_end_mask_0 = const()[name = string("op_1441_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1441_cast_fp16 = slice_by_index(begin = var_1441_begin_0, end = var_1441_end_0, end_mask = var_1441_end_mask_0, x = new_x_19_cast_fp16)[name = string("op_1441_cast_fp16")];
+            string x_125_pad_type_0 = const()[name = string("x_125_pad_type_0"), val = string("valid")];
+            int32 x_125_groups_0 = const()[name = string("x_125_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_125_strides_0 = const()[name = string("x_125_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_125_pad_0 = const()[name = string("x_125_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_125_dilations_0 = const()[name = string("x_125_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106004864))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106014144))))[name = string("encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_19_cast_fp16)[name = string("x_125_cast_fp16")];
+            tensor<int32, [3]> input_263_perm_0 = const()[name = string("input_263_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_127_axes_0 = const()[name = string("x_127_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106016256)))];
+            tensor<fp16, [1024]> encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106018368)))];
+            tensor<fp16, [1, 14, 1024]> input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = string("transpose_320")];
+            tensor<fp16, [1, 14, 1024]> x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = string("x_127_cast_fp16")];
+            tensor<int32, [3]> input_265_perm_0 = const()[name = string("input_265_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = string("transpose_319")];
+            tensor<fp16, [1, 1024, 14]> input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = string("input_267_cast_fp16")];
+            string x_129_pad_type_0 = const()[name = string("x_129_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_129_strides_0 = const()[name = string("x_129_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_129_pad_0 = const()[name = string("x_129_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_129_dilations_0 = const()[name = string("x_129_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_129_groups_0 = const()[name = string("x_129_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106020480))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107069120))))[name = string("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_129_cast_fp16 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("x_129_cast_fp16")];
+            tensor<int32, [3]> input_269_perm_0 = const()[name = string("input_269_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = string("transpose_318")];
+            tensor<fp16, [1, 14, 1024]> input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = string("input_271_cast_fp16")];
+            tensor<int32, [1]> input_273_axes_0 = const()[name = string("input_273_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107071232)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107073344)))];
+            tensor<fp16, [1, 14, 1024]> input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = string("input_273_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107075456))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110221248))))[name = string("encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_4_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110221440)))];
+            tensor<fp16, [1, 14, 4096]> linear_44_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized, x = input_273_cast_fp16)[name = string("linear_44_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = string("input_277_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110229696))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113375488))))[name = string("encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113375680)))];
+            tensor<fp16, [1, 14, 1024]> linear_45_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized, x = input_277_cast_fp16)[name = string("linear_45_cast_fp16")];
+            fp16 var_1484_to_fp16 = const()[name = string("op_1484_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1485_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1484_to_fp16)[name = string("op_1485_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1485_cast_fp16)[name = string("input_283_cast_fp16")];
+            tensor<int32, [1]> input_285_axes_0 = const()[name = string("input_285_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113377792)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113379904)))];
+            tensor<fp16, [1, 14, 1024]> input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = string("input_285_cast_fp16")];
+            tensor<int32, [4]> cache_21_begin_0 = const()[name = string("cache_21_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_21_end_0 = const()[name = string("cache_21_end_0"), val = tensor<int32, [4]>([6, 1, 42, 1024])];
+            tensor<bool, [4]> cache_21_end_mask_0 = const()[name = string("cache_21_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_21_squeeze_mask_0 = const()[name = string("cache_21_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_21_cast_fp16 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_21_cast_fp16")];
+            tensor<int32, [4]> cache_23_begin_0 = const()[name = string("cache_23_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_23_end_0 = const()[name = string("cache_23_end_0"), val = tensor<int32, [4]>([6, 1, 1024, 8])];
+            tensor<bool, [4]> cache_23_end_mask_0 = const()[name = string("cache_23_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_23_squeeze_mask_0 = const()[name = string("cache_23_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_23_cast_fp16 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_23_cast_fp16")];
+            tensor<int32, [1]> input_287_axes_0 = const()[name = string("input_287_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113382016)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113384128)))];
+            tensor<fp16, [1, 14, 1024]> input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = string("input_287_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113386240))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116532032))))[name = string("encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_5_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116532224)))];
+            tensor<fp16, [1, 14, 4096]> linear_46_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = string("linear_46_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("input_291_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116540480))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119686272))))[name = string("encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119686464)))];
+            tensor<fp16, [1, 14, 1024]> linear_47_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized, x = input_291_cast_fp16)[name = string("linear_47_cast_fp16")];
+            fp16 var_1521_to_fp16 = const()[name = string("op_1521_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1522_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1521_to_fp16)[name = string("op_1522_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1522_cast_fp16)[name = string("input_297_cast_fp16")];
+            tensor<int32, [1]> key_11_axes_0 = const()[name = string("key_11_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119688576)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119690688)))];
+            tensor<fp16, [1, 14, 1024]> key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = string("key_11_cast_fp16")];
+            bool input_299_interleave_0 = const()[name = string("input_299_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_299_cast_fp16 = concat(axis = var_69, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = string("input_299_cast_fp16")];
+            tensor<int32, [3]> var_1544_begin_0 = const()[name = string("op_1544_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_1544_end_0 = const()[name = string("op_1544_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1544_end_mask_0 = const()[name = string("op_1544_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_1544_cast_fp16 = slice_by_index(begin = var_1544_begin_0, end = var_1544_end_0, end_mask = var_1544_end_mask_0, x = cache_21_cast_fp16)[name = string("op_1544_cast_fp16")];
+            bool var_1550_interleave_0 = const()[name = string("op_1550_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1550_cast_fp16 = concat(axis = var_69, interleave = var_1550_interleave_0, values = (var_1544_cast_fp16, key_11_cast_fp16))[name = string("op_1550_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119692800))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120479296))))[name = string("encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120479488)))];
+            tensor<fp16, [1, 14, 1024]> linear_48_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized, x = key_11_cast_fp16)[name = string("linear_48_cast_fp16")];
+            tensor<int32, [4]> var_1555 = const()[name = string("op_1555"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_31_cast_fp16 = reshape(shape = var_1555, x = linear_48_cast_fp16)[name = string("q_31_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120481600))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121268096))))[name = string("encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121268288)))];
+            tensor<fp16, [1, 56, 1024]> linear_49_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_49_cast_fp16")];
+            tensor<int32, [4]> var_1560 = const()[name = string("op_1560"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_21_cast_fp16 = reshape(shape = var_1560, x = linear_49_cast_fp16)[name = string("k_21_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121270400))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122056896))))[name = string("encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122057088)))];
+            tensor<fp16, [1, 56, 1024]> linear_50_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_50_cast_fp16")];
+            tensor<int32, [4]> var_1565 = const()[name = string("op_1565"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_11_cast_fp16 = reshape(shape = var_1565, x = linear_50_cast_fp16)[name = string("v_11_cast_fp16")];
+            tensor<int32, [4]> value_19_perm_0 = const()[name = string("value_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122059200)))];
+            tensor<fp16, [1, 14, 8, 128]> var_1578_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = string("op_1578_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122061312)))];
+            tensor<fp16, [1, 14, 8, 128]> var_1580_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = string("op_1580_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_11_perm_0 = const()[name = string("q_with_bias_v_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_137_transpose_x_0 = const()[name = string("x_137_transpose_x_0"), val = bool(false)];
+            bool x_137_transpose_y_0 = const()[name = string("x_137_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_1582_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122063424))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122177152))))[name = string("op_1582_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1580_cast_fp16)[name = string("transpose_317")];
+            tensor<fp16, [1, 8, 14, 111]> x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = op_1582_to_fp16_quantized)[name = string("x_137_cast_fp16")];
+            tensor<int32, [8]> x_139_pad_0 = const()[name = string("x_139_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_139_mode_0 = const()[name = string("x_139_mode_0"), val = string("constant")];
+            fp16 const_144_to_fp16 = const()[name = string("const_144_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_139_cast_fp16 = pad(constant_val = const_144_to_fp16, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137_cast_fp16)[name = string("x_139_cast_fp16")];
+            tensor<int32, [4]> var_1590 = const()[name = string("op_1590"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_141_cast_fp16 = reshape(shape = var_1590, x = x_139_cast_fp16)[name = string("x_141_cast_fp16")];
+            tensor<int32, [4]> var_1594_begin_0 = const()[name = string("op_1594_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1594_end_0 = const()[name = string("op_1594_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_1594_end_mask_0 = const()[name = string("op_1594_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_1594_cast_fp16 = slice_by_index(begin = var_1594_begin_0, end = var_1594_end_0, end_mask = var_1594_end_mask_0, x = x_141_cast_fp16)[name = string("op_1594_cast_fp16")];
+            tensor<int32, [4]> var_1595 = const()[name = string("op_1595"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_21_cast_fp16 = reshape(shape = var_1595, x = var_1594_cast_fp16)[name = string("matrix_bd_21_cast_fp16")];
+            bool matrix_ac_11_transpose_x_0 = const()[name = string("matrix_ac_11_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_11_transpose_y_0 = const()[name = string("matrix_ac_11_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_107 = transpose(perm = transpose_107_perm_0, x = k_21_cast_fp16)[name = string("transpose_315")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_106 = transpose(perm = transpose_106_perm_0, x = var_1578_cast_fp16)[name = string("transpose_316")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_106, y = transpose_107)[name = string("matrix_ac_11_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_23_begin_0 = const()[name = string("matrix_bd_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_23_end_0 = const()[name = string("matrix_bd_23_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_23_end_mask_0 = const()[name = string("matrix_bd_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21_cast_fp16)[name = string("matrix_bd_23_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_1604_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = string("op_1604_cast_fp16")];
+            fp16 _inversed_scores_21_y_0_to_fp16 = const()[name = string("_inversed_scores_21_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_21_cast_fp16 = mul(x = var_1604_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = string("_inversed_scores_21_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_23_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_11)[name = string("scores_23_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_1610_cast_fp16 = softmax(axis = var_60, x = scores_23_cast_fp16)[name = string("op_1610_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_301_cast_fp16 = select(a = var_45_to_fp16, b = var_1610_cast_fp16, cond = mask_11)[name = string("input_301_cast_fp16")];
+            bool x_143_transpose_x_0 = const()[name = string("x_143_transpose_x_0"), val = bool(false)];
+            bool x_143_transpose_y_0 = const()[name = string("x_143_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_11_cast_fp16)[name = string("transpose_314")];
+            tensor<fp16, [1, 8, 14, 128]> x_143_cast_fp16 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301_cast_fp16, y = value_19_cast_fp16)[name = string("x_143_cast_fp16")];
+            tensor<int32, [4]> var_1614_perm_0 = const()[name = string("op_1614_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1615 = const()[name = string("op_1615"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_1614_cast_fp16 = transpose(perm = var_1614_perm_0, x = x_143_cast_fp16)[name = string("transpose_313")];
+            tensor<fp16, [1, 14, 1024]> input_303_cast_fp16 = reshape(shape = var_1615, x = var_1614_cast_fp16)[name = string("input_303_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122177472))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122963968))))[name = string("encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122964160)))];
+            tensor<fp16, [1, 14, 1024]> linear_52_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized, x = input_303_cast_fp16)[name = string("linear_52_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = string("input_307_cast_fp16")];
+            tensor<int32, [1]> x_147_axes_0 = const()[name = string("x_147_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122966272)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122968384)))];
+            tensor<fp16, [1, 14, 1024]> x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = string("x_147_cast_fp16")];
+            tensor<int32, [3]> input_309_perm_0 = const()[name = string("input_309_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_311_pad_type_0 = const()[name = string("input_311_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_311_strides_0 = const()[name = string("input_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_311_pad_0 = const()[name = string("input_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_311_dilations_0 = const()[name = string("input_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_311_groups_0 = const()[name = string("input_311_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122970496))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125067712))))[name = string("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = string("transpose_312")];
+            tensor<fp16, [1, 2048, 14]> input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_309_cast_fp16)[name = string("input_311_cast_fp16")];
+            int32 x_149_split_num_splits_0 = const()[name = string("x_149_split_num_splits_0"), val = int32(2)];
+            int32 x_149_split_axis_0 = const()[name = string("x_149_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_149_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_149_split_cast_fp16_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311_cast_fp16)[name = string("x_149_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = string("x_149_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = string("x_149_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_313_cast_fp16 = select(a = var_45_to_fp16, b = x_149_cast_fp16, cond = var_576)[name = string("input_313_cast_fp16")];
+            bool new_x_23_interleave_0 = const()[name = string("new_x_23_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_23_cast_fp16 = concat(axis = var_60, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = string("new_x_23_cast_fp16")];
+            tensor<int32, [3]> var_1654_begin_0 = const()[name = string("op_1654_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_1654_end_0 = const()[name = string("op_1654_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_1654_end_mask_0 = const()[name = string("op_1654_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1654_cast_fp16 = slice_by_index(begin = var_1654_begin_0, end = var_1654_end_0, end_mask = var_1654_end_mask_0, x = new_x_23_cast_fp16)[name = string("op_1654_cast_fp16")];
+            string x_151_pad_type_0 = const()[name = string("x_151_pad_type_0"), val = string("valid")];
+            int32 x_151_groups_0 = const()[name = string("x_151_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_151_strides_0 = const()[name = string("x_151_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_151_pad_0 = const()[name = string("x_151_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_151_dilations_0 = const()[name = string("x_151_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125071872))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125081152))))[name = string("encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_151_cast_fp16 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_23_cast_fp16)[name = string("x_151_cast_fp16")];
+            tensor<int32, [3]> input_315_perm_0 = const()[name = string("input_315_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_153_axes_0 = const()[name = string("x_153_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125083264)))];
+            tensor<fp16, [1024]> encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125085376)))];
+            tensor<fp16, [1, 14, 1024]> input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = string("transpose_311")];
+            tensor<fp16, [1, 14, 1024]> x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = string("x_153_cast_fp16")];
+            tensor<int32, [3]> input_317_perm_0 = const()[name = string("input_317_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = string("transpose_310")];
+            tensor<fp16, [1, 1024, 14]> input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = string("input_319_cast_fp16")];
+            string x_155_pad_type_0 = const()[name = string("x_155_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_155_strides_0 = const()[name = string("x_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_155_pad_0 = const()[name = string("x_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_155_dilations_0 = const()[name = string("x_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_155_groups_0 = const()[name = string("x_155_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125087488))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126136128))))[name = string("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_319_cast_fp16)[name = string("x_155_cast_fp16")];
+            tensor<int32, [3]> input_321_perm_0 = const()[name = string("input_321_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = string("transpose_309")];
+            tensor<fp16, [1, 14, 1024]> input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = string("input_323_cast_fp16")];
+            tensor<int32, [1]> input_325_axes_0 = const()[name = string("input_325_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126138240)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126140352)))];
+            tensor<fp16, [1, 14, 1024]> input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = string("input_325_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126142464))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129288256))))[name = string("encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_5_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129288448)))];
+            tensor<fp16, [1, 14, 4096]> linear_53_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized, x = input_325_cast_fp16)[name = string("linear_53_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("input_329_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129296704))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132442496))))[name = string("encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132442688)))];
+            tensor<fp16, [1, 14, 1024]> linear_54_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized, x = input_329_cast_fp16)[name = string("linear_54_cast_fp16")];
+            fp16 var_1697_to_fp16 = const()[name = string("op_1697_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1698_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1697_to_fp16)[name = string("op_1698_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1698_cast_fp16)[name = string("input_335_cast_fp16")];
+            tensor<int32, [1]> input_337_axes_0 = const()[name = string("input_337_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132444800)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132446912)))];
+            tensor<fp16, [1, 14, 1024]> input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = string("input_337_cast_fp16")];
+            tensor<int32, [4]> cache_25_begin_0 = const()[name = string("cache_25_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_25_end_0 = const()[name = string("cache_25_end_0"), val = tensor<int32, [4]>([7, 1, 42, 1024])];
+            tensor<bool, [4]> cache_25_end_mask_0 = const()[name = string("cache_25_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_25_squeeze_mask_0 = const()[name = string("cache_25_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_25_cast_fp16 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_25_cast_fp16")];
+            tensor<int32, [4]> cache_27_begin_0 = const()[name = string("cache_27_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_27_end_0 = const()[name = string("cache_27_end_0"), val = tensor<int32, [4]>([7, 1, 1024, 8])];
+            tensor<bool, [4]> cache_27_end_mask_0 = const()[name = string("cache_27_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_27_squeeze_mask_0 = const()[name = string("cache_27_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_27_cast_fp16 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_27_cast_fp16")];
+            tensor<int32, [1]> input_339_axes_0 = const()[name = string("input_339_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132449024)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132451136)))];
+            tensor<fp16, [1, 14, 1024]> input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = string("input_339_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132453248))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135599040))))[name = string("encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_6_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135599232)))];
+            tensor<fp16, [1, 14, 4096]> linear_55_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized, x = input_339_cast_fp16)[name = string("linear_55_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = string("input_343_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135607488))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138753280))))[name = string("encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138753472)))];
+            tensor<fp16, [1, 14, 1024]> linear_56_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized, x = input_343_cast_fp16)[name = string("linear_56_cast_fp16")];
+            fp16 var_1734_to_fp16 = const()[name = string("op_1734_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1735_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1734_to_fp16)[name = string("op_1735_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1735_cast_fp16)[name = string("input_349_cast_fp16")];
+            tensor<int32, [1]> key_13_axes_0 = const()[name = string("key_13_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138755584)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138757696)))];
+            tensor<fp16, [1, 14, 1024]> key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = string("key_13_cast_fp16")];
+            bool input_351_interleave_0 = const()[name = string("input_351_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_351_cast_fp16 = concat(axis = var_69, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = string("input_351_cast_fp16")];
+            tensor<int32, [3]> var_1757_begin_0 = const()[name = string("op_1757_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_1757_end_0 = const()[name = string("op_1757_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1757_end_mask_0 = const()[name = string("op_1757_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_1757_cast_fp16 = slice_by_index(begin = var_1757_begin_0, end = var_1757_end_0, end_mask = var_1757_end_mask_0, x = cache_25_cast_fp16)[name = string("op_1757_cast_fp16")];
+            bool var_1763_interleave_0 = const()[name = string("op_1763_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1763_cast_fp16 = concat(axis = var_69, interleave = var_1763_interleave_0, values = (var_1757_cast_fp16, key_13_cast_fp16))[name = string("op_1763_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138759808))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139546304))))[name = string("encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139546496)))];
+            tensor<fp16, [1, 14, 1024]> linear_57_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized, x = key_13_cast_fp16)[name = string("linear_57_cast_fp16")];
+            tensor<int32, [4]> var_1768 = const()[name = string("op_1768"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_37_cast_fp16 = reshape(shape = var_1768, x = linear_57_cast_fp16)[name = string("q_37_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139548608))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140335104))))[name = string("encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140335296)))];
+            tensor<fp16, [1, 56, 1024]> linear_58_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_58_cast_fp16")];
+            tensor<int32, [4]> var_1773 = const()[name = string("op_1773"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_25_cast_fp16 = reshape(shape = var_1773, x = linear_58_cast_fp16)[name = string("k_25_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140337408))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141123904))))[name = string("encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141124096)))];
+            tensor<fp16, [1, 56, 1024]> linear_59_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_59_cast_fp16")];
+            tensor<int32, [4]> var_1778 = const()[name = string("op_1778"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_13_cast_fp16 = reshape(shape = var_1778, x = linear_59_cast_fp16)[name = string("v_13_cast_fp16")];
+            tensor<int32, [4]> value_21_perm_0 = const()[name = string("value_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141126208)))];
+            tensor<fp16, [1, 14, 8, 128]> var_1791_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = string("op_1791_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141128320)))];
+            tensor<fp16, [1, 14, 8, 128]> var_1793_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = string("op_1793_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_13_perm_0 = const()[name = string("q_with_bias_v_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_163_transpose_x_0 = const()[name = string("x_163_transpose_x_0"), val = bool(false)];
+            bool x_163_transpose_y_0 = const()[name = string("x_163_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_1795_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141130432))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141244160))))[name = string("op_1795_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1793_cast_fp16)[name = string("transpose_308")];
+            tensor<fp16, [1, 8, 14, 111]> x_163_cast_fp16 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = op_1795_to_fp16_quantized)[name = string("x_163_cast_fp16")];
+            tensor<int32, [8]> x_165_pad_0 = const()[name = string("x_165_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_165_mode_0 = const()[name = string("x_165_mode_0"), val = string("constant")];
+            fp16 const_157_to_fp16 = const()[name = string("const_157_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_165_cast_fp16 = pad(constant_val = const_157_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = string("x_165_cast_fp16")];
+            tensor<int32, [4]> var_1803 = const()[name = string("op_1803"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_167_cast_fp16 = reshape(shape = var_1803, x = x_165_cast_fp16)[name = string("x_167_cast_fp16")];
+            tensor<int32, [4]> var_1807_begin_0 = const()[name = string("op_1807_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1807_end_0 = const()[name = string("op_1807_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_1807_end_mask_0 = const()[name = string("op_1807_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_1807_cast_fp16 = slice_by_index(begin = var_1807_begin_0, end = var_1807_end_0, end_mask = var_1807_end_mask_0, x = x_167_cast_fp16)[name = string("op_1807_cast_fp16")];
+            tensor<int32, [4]> var_1808 = const()[name = string("op_1808"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_25_cast_fp16 = reshape(shape = var_1808, x = var_1807_cast_fp16)[name = string("matrix_bd_25_cast_fp16")];
+            bool matrix_ac_13_transpose_x_0 = const()[name = string("matrix_ac_13_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_13_transpose_y_0 = const()[name = string("matrix_ac_13_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_109 = transpose(perm = transpose_109_perm_0, x = k_25_cast_fp16)[name = string("transpose_306")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1791_cast_fp16)[name = string("transpose_307")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_108, y = transpose_109)[name = string("matrix_ac_13_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_27_begin_0 = const()[name = string("matrix_bd_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_27_end_0 = const()[name = string("matrix_bd_27_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_27_end_mask_0 = const()[name = string("matrix_bd_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25_cast_fp16)[name = string("matrix_bd_27_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_1817_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = string("op_1817_cast_fp16")];
+            fp16 _inversed_scores_25_y_0_to_fp16 = const()[name = string("_inversed_scores_25_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_25_cast_fp16 = mul(x = var_1817_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = string("_inversed_scores_25_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_27_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_11)[name = string("scores_27_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_1823_cast_fp16 = softmax(axis = var_60, x = scores_27_cast_fp16)[name = string("op_1823_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_353_cast_fp16 = select(a = var_45_to_fp16, b = var_1823_cast_fp16, cond = mask_11)[name = string("input_353_cast_fp16")];
+            bool x_169_transpose_x_0 = const()[name = string("x_169_transpose_x_0"), val = bool(false)];
+            bool x_169_transpose_y_0 = const()[name = string("x_169_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_13_cast_fp16)[name = string("transpose_305")];
+            tensor<fp16, [1, 8, 14, 128]> x_169_cast_fp16 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353_cast_fp16, y = value_21_cast_fp16)[name = string("x_169_cast_fp16")];
+            tensor<int32, [4]> var_1827_perm_0 = const()[name = string("op_1827_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1828 = const()[name = string("op_1828"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_1827_cast_fp16 = transpose(perm = var_1827_perm_0, x = x_169_cast_fp16)[name = string("transpose_304")];
+            tensor<fp16, [1, 14, 1024]> input_355_cast_fp16 = reshape(shape = var_1828, x = var_1827_cast_fp16)[name = string("input_355_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141244480))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142030976))))[name = string("encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142031168)))];
+            tensor<fp16, [1, 14, 1024]> linear_61_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized, x = input_355_cast_fp16)[name = string("linear_61_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = string("input_359_cast_fp16")];
+            tensor<int32, [1]> x_173_axes_0 = const()[name = string("x_173_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142033280)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142035392)))];
+            tensor<fp16, [1, 14, 1024]> x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = string("x_173_cast_fp16")];
+            tensor<int32, [3]> input_361_perm_0 = const()[name = string("input_361_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_363_pad_type_0 = const()[name = string("input_363_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_363_strides_0 = const()[name = string("input_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_363_pad_0 = const()[name = string("input_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_363_dilations_0 = const()[name = string("input_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_363_groups_0 = const()[name = string("input_363_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142037504))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144134720))))[name = string("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = string("transpose_303")];
+            tensor<fp16, [1, 2048, 14]> input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_361_cast_fp16)[name = string("input_363_cast_fp16")];
+            int32 x_175_split_num_splits_0 = const()[name = string("x_175_split_num_splits_0"), val = int32(2)];
+            int32 x_175_split_axis_0 = const()[name = string("x_175_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_175_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_175_split_cast_fp16_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363_cast_fp16)[name = string("x_175_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = string("x_175_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = string("x_175_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_365_cast_fp16 = select(a = var_45_to_fp16, b = x_175_cast_fp16, cond = var_576)[name = string("input_365_cast_fp16")];
+            bool new_x_27_interleave_0 = const()[name = string("new_x_27_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_27_cast_fp16 = concat(axis = var_60, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = string("new_x_27_cast_fp16")];
+            tensor<int32, [3]> var_1867_begin_0 = const()[name = string("op_1867_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_1867_end_0 = const()[name = string("op_1867_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_1867_end_mask_0 = const()[name = string("op_1867_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1867_cast_fp16 = slice_by_index(begin = var_1867_begin_0, end = var_1867_end_0, end_mask = var_1867_end_mask_0, x = new_x_27_cast_fp16)[name = string("op_1867_cast_fp16")];
+            string x_177_pad_type_0 = const()[name = string("x_177_pad_type_0"), val = string("valid")];
+            int32 x_177_groups_0 = const()[name = string("x_177_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_177_strides_0 = const()[name = string("x_177_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_177_pad_0 = const()[name = string("x_177_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_177_dilations_0 = const()[name = string("x_177_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144138880))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144148160))))[name = string("encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_177_cast_fp16 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_27_cast_fp16)[name = string("x_177_cast_fp16")];
+            tensor<int32, [3]> input_367_perm_0 = const()[name = string("input_367_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_179_axes_0 = const()[name = string("x_179_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144150272)))];
+            tensor<fp16, [1024]> encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144152384)))];
+            tensor<fp16, [1, 14, 1024]> input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = string("transpose_302")];
+            tensor<fp16, [1, 14, 1024]> x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = string("x_179_cast_fp16")];
+            tensor<int32, [3]> input_369_perm_0 = const()[name = string("input_369_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = string("transpose_301")];
+            tensor<fp16, [1, 1024, 14]> input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = string("input_371_cast_fp16")];
+            string x_181_pad_type_0 = const()[name = string("x_181_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_181_strides_0 = const()[name = string("x_181_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_181_pad_0 = const()[name = string("x_181_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_181_dilations_0 = const()[name = string("x_181_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_181_groups_0 = const()[name = string("x_181_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144154496))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145203136))))[name = string("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_181_cast_fp16 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_371_cast_fp16)[name = string("x_181_cast_fp16")];
+            tensor<int32, [3]> input_373_perm_0 = const()[name = string("input_373_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = string("transpose_300")];
+            tensor<fp16, [1, 14, 1024]> input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = string("input_375_cast_fp16")];
+            tensor<int32, [1]> input_377_axes_0 = const()[name = string("input_377_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145205248)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145207360)))];
+            tensor<fp16, [1, 14, 1024]> input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = string("input_377_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145209472))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148355264))))[name = string("encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_6_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148355456)))];
+            tensor<fp16, [1, 14, 4096]> linear_62_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized, x = input_377_cast_fp16)[name = string("linear_62_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = string("input_381_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148363712))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151509504))))[name = string("encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151509696)))];
+            tensor<fp16, [1, 14, 1024]> linear_63_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized, x = input_381_cast_fp16)[name = string("linear_63_cast_fp16")];
+            fp16 var_1910_to_fp16 = const()[name = string("op_1910_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1911_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1910_to_fp16)[name = string("op_1911_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1911_cast_fp16)[name = string("input_387_cast_fp16")];
+            tensor<int32, [1]> input_389_axes_0 = const()[name = string("input_389_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151511808)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151513920)))];
+            tensor<fp16, [1, 14, 1024]> input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = string("input_389_cast_fp16")];
+            tensor<int32, [4]> cache_29_begin_0 = const()[name = string("cache_29_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_29_end_0 = const()[name = string("cache_29_end_0"), val = tensor<int32, [4]>([8, 1, 42, 1024])];
+            tensor<bool, [4]> cache_29_end_mask_0 = const()[name = string("cache_29_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_29_squeeze_mask_0 = const()[name = string("cache_29_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_29_cast_fp16 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_29_cast_fp16")];
+            tensor<int32, [4]> cache_31_begin_0 = const()[name = string("cache_31_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_31_end_0 = const()[name = string("cache_31_end_0"), val = tensor<int32, [4]>([8, 1, 1024, 8])];
+            tensor<bool, [4]> cache_31_end_mask_0 = const()[name = string("cache_31_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_31_squeeze_mask_0 = const()[name = string("cache_31_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_31_cast_fp16 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_31_cast_fp16")];
+            tensor<int32, [1]> input_391_axes_0 = const()[name = string("input_391_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151516032)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151518144)))];
+            tensor<fp16, [1, 14, 1024]> input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = string("input_391_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151520256))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154666048))))[name = string("encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_7_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154666240)))];
+            tensor<fp16, [1, 14, 4096]> linear_64_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized, x = input_391_cast_fp16)[name = string("linear_64_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = string("input_395_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154674496))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157820288))))[name = string("encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157820480)))];
+            tensor<fp16, [1, 14, 1024]> linear_65_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized, x = input_395_cast_fp16)[name = string("linear_65_cast_fp16")];
+            fp16 var_1947_to_fp16 = const()[name = string("op_1947_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_1948_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1947_to_fp16)[name = string("op_1948_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1948_cast_fp16)[name = string("input_401_cast_fp16")];
+            tensor<int32, [1]> key_15_axes_0 = const()[name = string("key_15_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157822592)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157824704)))];
+            tensor<fp16, [1, 14, 1024]> key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = string("key_15_cast_fp16")];
+            bool input_403_interleave_0 = const()[name = string("input_403_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_403_cast_fp16 = concat(axis = var_69, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = string("input_403_cast_fp16")];
+            tensor<int32, [3]> var_1970_begin_0 = const()[name = string("op_1970_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_1970_end_0 = const()[name = string("op_1970_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1970_end_mask_0 = const()[name = string("op_1970_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_1970_cast_fp16 = slice_by_index(begin = var_1970_begin_0, end = var_1970_end_0, end_mask = var_1970_end_mask_0, x = cache_29_cast_fp16)[name = string("op_1970_cast_fp16")];
+            bool var_1976_interleave_0 = const()[name = string("op_1976_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1976_cast_fp16 = concat(axis = var_69, interleave = var_1976_interleave_0, values = (var_1970_cast_fp16, key_15_cast_fp16))[name = string("op_1976_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157826816))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158613312))))[name = string("encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158613504)))];
+            tensor<fp16, [1, 14, 1024]> linear_66_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized, x = key_15_cast_fp16)[name = string("linear_66_cast_fp16")];
+            tensor<int32, [4]> var_1981 = const()[name = string("op_1981"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_43_cast_fp16 = reshape(shape = var_1981, x = linear_66_cast_fp16)[name = string("q_43_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158615616))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159402112))))[name = string("encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159402304)))];
+            tensor<fp16, [1, 56, 1024]> linear_67_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_67_cast_fp16")];
+            tensor<int32, [4]> var_1986 = const()[name = string("op_1986"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_29_cast_fp16 = reshape(shape = var_1986, x = linear_67_cast_fp16)[name = string("k_29_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159404416))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160190912))))[name = string("encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160191104)))];
+            tensor<fp16, [1, 56, 1024]> linear_68_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_68_cast_fp16")];
+            tensor<int32, [4]> var_1991 = const()[name = string("op_1991"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_15_cast_fp16 = reshape(shape = var_1991, x = linear_68_cast_fp16)[name = string("v_15_cast_fp16")];
+            tensor<int32, [4]> value_23_perm_0 = const()[name = string("value_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160193216)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2004_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = string("op_2004_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160195328)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2006_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = string("op_2006_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_15_perm_0 = const()[name = string("q_with_bias_v_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_189_transpose_x_0 = const()[name = string("x_189_transpose_x_0"), val = bool(false)];
+            bool x_189_transpose_y_0 = const()[name = string("x_189_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_2008_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160197440))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160311168))))[name = string("op_2008_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_2006_cast_fp16)[name = string("transpose_299")];
+            tensor<fp16, [1, 8, 14, 111]> x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = op_2008_to_fp16_quantized)[name = string("x_189_cast_fp16")];
+            tensor<int32, [8]> x_191_pad_0 = const()[name = string("x_191_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_191_mode_0 = const()[name = string("x_191_mode_0"), val = string("constant")];
+            fp16 const_170_to_fp16 = const()[name = string("const_170_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_191_cast_fp16 = pad(constant_val = const_170_to_fp16, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189_cast_fp16)[name = string("x_191_cast_fp16")];
+            tensor<int32, [4]> var_2016 = const()[name = string("op_2016"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_193_cast_fp16 = reshape(shape = var_2016, x = x_191_cast_fp16)[name = string("x_193_cast_fp16")];
+            tensor<int32, [4]> var_2020_begin_0 = const()[name = string("op_2020_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2020_end_0 = const()[name = string("op_2020_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_2020_end_mask_0 = const()[name = string("op_2020_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_2020_cast_fp16 = slice_by_index(begin = var_2020_begin_0, end = var_2020_end_0, end_mask = var_2020_end_mask_0, x = x_193_cast_fp16)[name = string("op_2020_cast_fp16")];
+            tensor<int32, [4]> var_2021 = const()[name = string("op_2021"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_29_cast_fp16 = reshape(shape = var_2021, x = var_2020_cast_fp16)[name = string("matrix_bd_29_cast_fp16")];
+            bool matrix_ac_15_transpose_x_0 = const()[name = string("matrix_ac_15_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_15_transpose_y_0 = const()[name = string("matrix_ac_15_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_111 = transpose(perm = transpose_111_perm_0, x = k_29_cast_fp16)[name = string("transpose_297")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_110 = transpose(perm = transpose_110_perm_0, x = var_2004_cast_fp16)[name = string("transpose_298")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_110, y = transpose_111)[name = string("matrix_ac_15_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_31_begin_0 = const()[name = string("matrix_bd_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_31_end_0 = const()[name = string("matrix_bd_31_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_31_end_mask_0 = const()[name = string("matrix_bd_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29_cast_fp16)[name = string("matrix_bd_31_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2030_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = string("op_2030_cast_fp16")];
+            fp16 _inversed_scores_29_y_0_to_fp16 = const()[name = string("_inversed_scores_29_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_29_cast_fp16 = mul(x = var_2030_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = string("_inversed_scores_29_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_31_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_11)[name = string("scores_31_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2036_cast_fp16 = softmax(axis = var_60, x = scores_31_cast_fp16)[name = string("op_2036_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_405_cast_fp16 = select(a = var_45_to_fp16, b = var_2036_cast_fp16, cond = mask_11)[name = string("input_405_cast_fp16")];
+            bool x_195_transpose_x_0 = const()[name = string("x_195_transpose_x_0"), val = bool(false)];
+            bool x_195_transpose_y_0 = const()[name = string("x_195_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_15_cast_fp16)[name = string("transpose_296")];
+            tensor<fp16, [1, 8, 14, 128]> x_195_cast_fp16 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405_cast_fp16, y = value_23_cast_fp16)[name = string("x_195_cast_fp16")];
+            tensor<int32, [4]> var_2040_perm_0 = const()[name = string("op_2040_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2041 = const()[name = string("op_2041"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_2040_cast_fp16 = transpose(perm = var_2040_perm_0, x = x_195_cast_fp16)[name = string("transpose_295")];
+            tensor<fp16, [1, 14, 1024]> input_407_cast_fp16 = reshape(shape = var_2041, x = var_2040_cast_fp16)[name = string("input_407_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160311488))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161097984))))[name = string("encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161098176)))];
+            tensor<fp16, [1, 14, 1024]> linear_70_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = string("linear_70_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = string("input_411_cast_fp16")];
+            tensor<int32, [1]> x_199_axes_0 = const()[name = string("x_199_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161100288)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161102400)))];
+            tensor<fp16, [1, 14, 1024]> x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = string("x_199_cast_fp16")];
+            tensor<int32, [3]> input_413_perm_0 = const()[name = string("input_413_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_415_pad_type_0 = const()[name = string("input_415_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_415_strides_0 = const()[name = string("input_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_415_pad_0 = const()[name = string("input_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_415_dilations_0 = const()[name = string("input_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_415_groups_0 = const()[name = string("input_415_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161104512))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163201728))))[name = string("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = string("transpose_294")];
+            tensor<fp16, [1, 2048, 14]> input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_413_cast_fp16)[name = string("input_415_cast_fp16")];
+            int32 x_201_split_num_splits_0 = const()[name = string("x_201_split_num_splits_0"), val = int32(2)];
+            int32 x_201_split_axis_0 = const()[name = string("x_201_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_201_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_201_split_cast_fp16_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415_cast_fp16)[name = string("x_201_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = string("x_201_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = string("x_201_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_417_cast_fp16 = select(a = var_45_to_fp16, b = x_201_cast_fp16, cond = var_576)[name = string("input_417_cast_fp16")];
+            bool new_x_31_interleave_0 = const()[name = string("new_x_31_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_31_cast_fp16 = concat(axis = var_60, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = string("new_x_31_cast_fp16")];
+            tensor<int32, [3]> var_2080_begin_0 = const()[name = string("op_2080_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_2080_end_0 = const()[name = string("op_2080_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_2080_end_mask_0 = const()[name = string("op_2080_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2080_cast_fp16 = slice_by_index(begin = var_2080_begin_0, end = var_2080_end_0, end_mask = var_2080_end_mask_0, x = new_x_31_cast_fp16)[name = string("op_2080_cast_fp16")];
+            string x_203_pad_type_0 = const()[name = string("x_203_pad_type_0"), val = string("valid")];
+            int32 x_203_groups_0 = const()[name = string("x_203_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_203_strides_0 = const()[name = string("x_203_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_203_pad_0 = const()[name = string("x_203_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_203_dilations_0 = const()[name = string("x_203_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163205888))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163215168))))[name = string("encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_31_cast_fp16)[name = string("x_203_cast_fp16")];
+            tensor<int32, [3]> input_419_perm_0 = const()[name = string("input_419_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_205_axes_0 = const()[name = string("x_205_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163217280)))];
+            tensor<fp16, [1024]> encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163219392)))];
+            tensor<fp16, [1, 14, 1024]> input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = string("transpose_293")];
+            tensor<fp16, [1, 14, 1024]> x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = string("x_205_cast_fp16")];
+            tensor<int32, [3]> input_421_perm_0 = const()[name = string("input_421_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = string("transpose_292")];
+            tensor<fp16, [1, 1024, 14]> input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = string("input_423_cast_fp16")];
+            string x_207_pad_type_0 = const()[name = string("x_207_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_207_strides_0 = const()[name = string("x_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_207_pad_0 = const()[name = string("x_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_207_dilations_0 = const()[name = string("x_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_207_groups_0 = const()[name = string("x_207_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163221504))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164270144))))[name = string("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_207_cast_fp16 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_423_cast_fp16)[name = string("x_207_cast_fp16")];
+            tensor<int32, [3]> input_425_perm_0 = const()[name = string("input_425_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = string("transpose_291")];
+            tensor<fp16, [1, 14, 1024]> input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = string("input_427_cast_fp16")];
+            tensor<int32, [1]> input_429_axes_0 = const()[name = string("input_429_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164272256)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164274368)))];
+            tensor<fp16, [1, 14, 1024]> input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = string("input_429_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164276480))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167422272))))[name = string("encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_7_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167422464)))];
+            tensor<fp16, [1, 14, 4096]> linear_71_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized, x = input_429_cast_fp16)[name = string("linear_71_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = string("input_433_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167430720))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170576512))))[name = string("encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170576704)))];
+            tensor<fp16, [1, 14, 1024]> linear_72_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized, x = input_433_cast_fp16)[name = string("linear_72_cast_fp16")];
+            fp16 var_2123_to_fp16 = const()[name = string("op_2123_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_2124_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_2123_to_fp16)[name = string("op_2124_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_2124_cast_fp16)[name = string("input_439_cast_fp16")];
+            tensor<int32, [1]> input_441_axes_0 = const()[name = string("input_441_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170578816)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170580928)))];
+            tensor<fp16, [1, 14, 1024]> input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = string("input_441_cast_fp16")];
+            tensor<int32, [4]> cache_33_begin_0 = const()[name = string("cache_33_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_33_end_0 = const()[name = string("cache_33_end_0"), val = tensor<int32, [4]>([9, 1, 42, 1024])];
+            tensor<bool, [4]> cache_33_end_mask_0 = const()[name = string("cache_33_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_33_squeeze_mask_0 = const()[name = string("cache_33_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_33_cast_fp16 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_33_cast_fp16")];
+            tensor<int32, [4]> cache_35_begin_0 = const()[name = string("cache_35_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_35_end_0 = const()[name = string("cache_35_end_0"), val = tensor<int32, [4]>([9, 1, 1024, 8])];
+            tensor<bool, [4]> cache_35_end_mask_0 = const()[name = string("cache_35_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_35_squeeze_mask_0 = const()[name = string("cache_35_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_35_cast_fp16 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_35_cast_fp16")];
+            tensor<int32, [1]> input_443_axes_0 = const()[name = string("input_443_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170583040)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170585152)))];
+            tensor<fp16, [1, 14, 1024]> input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = string("input_443_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170587264))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173733056))))[name = string("encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_8_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173733248)))];
+            tensor<fp16, [1, 14, 4096]> linear_73_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("linear_73_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = string("input_447_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173741504))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176887296))))[name = string("encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176887488)))];
+            tensor<fp16, [1, 14, 1024]> linear_74_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized, x = input_447_cast_fp16)[name = string("linear_74_cast_fp16")];
+            fp16 var_2160_to_fp16 = const()[name = string("op_2160_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_2161_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_2160_to_fp16)[name = string("op_2161_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_2161_cast_fp16)[name = string("input_453_cast_fp16")];
+            tensor<int32, [1]> key_17_axes_0 = const()[name = string("key_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176889600)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176891712)))];
+            tensor<fp16, [1, 14, 1024]> key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = string("key_17_cast_fp16")];
+            bool input_455_interleave_0 = const()[name = string("input_455_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_455_cast_fp16 = concat(axis = var_69, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = string("input_455_cast_fp16")];
+            tensor<int32, [3]> var_2183_begin_0 = const()[name = string("op_2183_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_2183_end_0 = const()[name = string("op_2183_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2183_end_mask_0 = const()[name = string("op_2183_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_2183_cast_fp16 = slice_by_index(begin = var_2183_begin_0, end = var_2183_end_0, end_mask = var_2183_end_mask_0, x = cache_33_cast_fp16)[name = string("op_2183_cast_fp16")];
+            bool var_2189_interleave_0 = const()[name = string("op_2189_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_2189_cast_fp16 = concat(axis = var_69, interleave = var_2189_interleave_0, values = (var_2183_cast_fp16, key_17_cast_fp16))[name = string("op_2189_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176893824))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177680320))))[name = string("encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177680512)))];
+            tensor<fp16, [1, 14, 1024]> linear_75_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized, x = key_17_cast_fp16)[name = string("linear_75_cast_fp16")];
+            tensor<int32, [4]> var_2194 = const()[name = string("op_2194"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_49_cast_fp16 = reshape(shape = var_2194, x = linear_75_cast_fp16)[name = string("q_49_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177682624))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178469120))))[name = string("encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178469312)))];
+            tensor<fp16, [1, 56, 1024]> linear_76_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_76_cast_fp16")];
+            tensor<int32, [4]> var_2199 = const()[name = string("op_2199"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_33_cast_fp16 = reshape(shape = var_2199, x = linear_76_cast_fp16)[name = string("k_33_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178471424))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179257920))))[name = string("encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179258112)))];
+            tensor<fp16, [1, 56, 1024]> linear_77_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_77_cast_fp16")];
+            tensor<int32, [4]> var_2204 = const()[name = string("op_2204"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_17_cast_fp16 = reshape(shape = var_2204, x = linear_77_cast_fp16)[name = string("v_17_cast_fp16")];
+            tensor<int32, [4]> value_25_perm_0 = const()[name = string("value_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179260224)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2217_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = string("op_2217_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179262336)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2219_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = string("op_2219_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_17_perm_0 = const()[name = string("q_with_bias_v_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_215_transpose_x_0 = const()[name = string("x_215_transpose_x_0"), val = bool(false)];
+            bool x_215_transpose_y_0 = const()[name = string("x_215_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_2221_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179264448))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179378176))))[name = string("op_2221_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2219_cast_fp16)[name = string("transpose_290")];
+            tensor<fp16, [1, 8, 14, 111]> x_215_cast_fp16 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = op_2221_to_fp16_quantized)[name = string("x_215_cast_fp16")];
+            tensor<int32, [8]> x_217_pad_0 = const()[name = string("x_217_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_217_mode_0 = const()[name = string("x_217_mode_0"), val = string("constant")];
+            fp16 const_183_to_fp16 = const()[name = string("const_183_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_217_cast_fp16 = pad(constant_val = const_183_to_fp16, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215_cast_fp16)[name = string("x_217_cast_fp16")];
+            tensor<int32, [4]> var_2229 = const()[name = string("op_2229"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_219_cast_fp16 = reshape(shape = var_2229, x = x_217_cast_fp16)[name = string("x_219_cast_fp16")];
+            tensor<int32, [4]> var_2233_begin_0 = const()[name = string("op_2233_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2233_end_0 = const()[name = string("op_2233_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_2233_end_mask_0 = const()[name = string("op_2233_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_2233_cast_fp16 = slice_by_index(begin = var_2233_begin_0, end = var_2233_end_0, end_mask = var_2233_end_mask_0, x = x_219_cast_fp16)[name = string("op_2233_cast_fp16")];
+            tensor<int32, [4]> var_2234 = const()[name = string("op_2234"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_33_cast_fp16 = reshape(shape = var_2234, x = var_2233_cast_fp16)[name = string("matrix_bd_33_cast_fp16")];
+            bool matrix_ac_17_transpose_x_0 = const()[name = string("matrix_ac_17_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_17_transpose_y_0 = const()[name = string("matrix_ac_17_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_113 = transpose(perm = transpose_113_perm_0, x = k_33_cast_fp16)[name = string("transpose_288")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_112 = transpose(perm = transpose_112_perm_0, x = var_2217_cast_fp16)[name = string("transpose_289")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_112, y = transpose_113)[name = string("matrix_ac_17_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_35_begin_0 = const()[name = string("matrix_bd_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_35_end_0 = const()[name = string("matrix_bd_35_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_35_end_mask_0 = const()[name = string("matrix_bd_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33_cast_fp16)[name = string("matrix_bd_35_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2243_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = string("op_2243_cast_fp16")];
+            fp16 _inversed_scores_33_y_0_to_fp16 = const()[name = string("_inversed_scores_33_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_33_cast_fp16 = mul(x = var_2243_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = string("_inversed_scores_33_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_35_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_11)[name = string("scores_35_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2249_cast_fp16 = softmax(axis = var_60, x = scores_35_cast_fp16)[name = string("op_2249_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_457_cast_fp16 = select(a = var_45_to_fp16, b = var_2249_cast_fp16, cond = mask_11)[name = string("input_457_cast_fp16")];
+            bool x_221_transpose_x_0 = const()[name = string("x_221_transpose_x_0"), val = bool(false)];
+            bool x_221_transpose_y_0 = const()[name = string("x_221_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_17_cast_fp16)[name = string("transpose_287")];
+            tensor<fp16, [1, 8, 14, 128]> x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457_cast_fp16, y = value_25_cast_fp16)[name = string("x_221_cast_fp16")];
+            tensor<int32, [4]> var_2253_perm_0 = const()[name = string("op_2253_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2254 = const()[name = string("op_2254"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_2253_cast_fp16 = transpose(perm = var_2253_perm_0, x = x_221_cast_fp16)[name = string("transpose_286")];
+            tensor<fp16, [1, 14, 1024]> input_459_cast_fp16 = reshape(shape = var_2254, x = var_2253_cast_fp16)[name = string("input_459_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179378496))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180164992))))[name = string("encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180165184)))];
+            tensor<fp16, [1, 14, 1024]> linear_79_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized, x = input_459_cast_fp16)[name = string("linear_79_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = string("input_463_cast_fp16")];
+            tensor<int32, [1]> x_225_axes_0 = const()[name = string("x_225_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180167296)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180169408)))];
+            tensor<fp16, [1, 14, 1024]> x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = string("x_225_cast_fp16")];
+            tensor<int32, [3]> input_465_perm_0 = const()[name = string("input_465_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_467_pad_type_0 = const()[name = string("input_467_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_467_strides_0 = const()[name = string("input_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_467_pad_0 = const()[name = string("input_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_467_dilations_0 = const()[name = string("input_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_467_groups_0 = const()[name = string("input_467_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180171520))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182268736))))[name = string("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = string("transpose_285")];
+            tensor<fp16, [1, 2048, 14]> input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_465_cast_fp16)[name = string("input_467_cast_fp16")];
+            int32 x_227_split_num_splits_0 = const()[name = string("x_227_split_num_splits_0"), val = int32(2)];
+            int32 x_227_split_axis_0 = const()[name = string("x_227_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_227_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_227_split_cast_fp16_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467_cast_fp16)[name = string("x_227_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = string("x_227_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = string("x_227_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_469_cast_fp16 = select(a = var_45_to_fp16, b = x_227_cast_fp16, cond = var_576)[name = string("input_469_cast_fp16")];
+            bool new_x_35_interleave_0 = const()[name = string("new_x_35_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_35_cast_fp16 = concat(axis = var_60, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = string("new_x_35_cast_fp16")];
+            tensor<int32, [3]> var_2293_begin_0 = const()[name = string("op_2293_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_2293_end_0 = const()[name = string("op_2293_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_2293_end_mask_0 = const()[name = string("op_2293_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2293_cast_fp16 = slice_by_index(begin = var_2293_begin_0, end = var_2293_end_0, end_mask = var_2293_end_mask_0, x = new_x_35_cast_fp16)[name = string("op_2293_cast_fp16")];
+            string x_229_pad_type_0 = const()[name = string("x_229_pad_type_0"), val = string("valid")];
+            int32 x_229_groups_0 = const()[name = string("x_229_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_229_strides_0 = const()[name = string("x_229_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_229_pad_0 = const()[name = string("x_229_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_229_dilations_0 = const()[name = string("x_229_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182272896))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182282176))))[name = string("encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_229_cast_fp16 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_35_cast_fp16)[name = string("x_229_cast_fp16")];
+            tensor<int32, [3]> input_471_perm_0 = const()[name = string("input_471_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_231_axes_0 = const()[name = string("x_231_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182284288)))];
+            tensor<fp16, [1024]> encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182286400)))];
+            tensor<fp16, [1, 14, 1024]> input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = string("transpose_284")];
+            tensor<fp16, [1, 14, 1024]> x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = string("x_231_cast_fp16")];
+            tensor<int32, [3]> input_473_perm_0 = const()[name = string("input_473_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = string("transpose_283")];
+            tensor<fp16, [1, 1024, 14]> input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = string("input_475_cast_fp16")];
+            string x_233_pad_type_0 = const()[name = string("x_233_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_233_strides_0 = const()[name = string("x_233_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_233_pad_0 = const()[name = string("x_233_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_233_dilations_0 = const()[name = string("x_233_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_233_groups_0 = const()[name = string("x_233_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182288512))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183337152))))[name = string("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_233_cast_fp16 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_475_cast_fp16)[name = string("x_233_cast_fp16")];
+            tensor<int32, [3]> input_477_perm_0 = const()[name = string("input_477_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = string("transpose_282")];
+            tensor<fp16, [1, 14, 1024]> input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = string("input_479_cast_fp16")];
+            tensor<int32, [1]> input_481_axes_0 = const()[name = string("input_481_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183339264)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183341376)))];
+            tensor<fp16, [1, 14, 1024]> input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = string("input_481_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183343488))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186489280))))[name = string("encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_8_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186489472)))];
+            tensor<fp16, [1, 14, 4096]> linear_80_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized, x = input_481_cast_fp16)[name = string("linear_80_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = string("input_485_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186497728))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189643520))))[name = string("encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189643712)))];
+            tensor<fp16, [1, 14, 1024]> linear_81_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized, x = input_485_cast_fp16)[name = string("linear_81_cast_fp16")];
+            fp16 var_2336_to_fp16 = const()[name = string("op_2336_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_2337_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2336_to_fp16)[name = string("op_2337_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2337_cast_fp16)[name = string("input_491_cast_fp16")];
+            tensor<int32, [1]> input_493_axes_0 = const()[name = string("input_493_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189645824)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189647936)))];
+            tensor<fp16, [1, 14, 1024]> input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = string("input_493_cast_fp16")];
+            tensor<int32, [4]> cache_37_begin_0 = const()[name = string("cache_37_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_37_end_0 = const()[name = string("cache_37_end_0"), val = tensor<int32, [4]>([10, 1, 42, 1024])];
+            tensor<bool, [4]> cache_37_end_mask_0 = const()[name = string("cache_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_37_squeeze_mask_0 = const()[name = string("cache_37_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_37_cast_fp16 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_37_cast_fp16")];
+            tensor<int32, [4]> cache_39_begin_0 = const()[name = string("cache_39_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_39_end_0 = const()[name = string("cache_39_end_0"), val = tensor<int32, [4]>([10, 1, 1024, 8])];
+            tensor<bool, [4]> cache_39_end_mask_0 = const()[name = string("cache_39_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_39_squeeze_mask_0 = const()[name = string("cache_39_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_39_cast_fp16 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_39_cast_fp16")];
+            tensor<int32, [1]> input_495_axes_0 = const()[name = string("input_495_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189650048)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189652160)))];
+            tensor<fp16, [1, 14, 1024]> input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = string("input_495_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189654272))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192800064))))[name = string("encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_9_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192800256)))];
+            tensor<fp16, [1, 14, 4096]> linear_82_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized, x = input_495_cast_fp16)[name = string("linear_82_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = string("input_499_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192808512))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195954304))))[name = string("encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195954496)))];
+            tensor<fp16, [1, 14, 1024]> linear_83_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized, x = input_499_cast_fp16)[name = string("linear_83_cast_fp16")];
+            fp16 var_2373_to_fp16 = const()[name = string("op_2373_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_2374_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2373_to_fp16)[name = string("op_2374_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2374_cast_fp16)[name = string("input_505_cast_fp16")];
+            tensor<int32, [1]> key_19_axes_0 = const()[name = string("key_19_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195956608)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195958720)))];
+            tensor<fp16, [1, 14, 1024]> key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = string("key_19_cast_fp16")];
+            bool input_507_interleave_0 = const()[name = string("input_507_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_507_cast_fp16 = concat(axis = var_69, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = string("input_507_cast_fp16")];
+            tensor<int32, [3]> var_2396_begin_0 = const()[name = string("op_2396_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_2396_end_0 = const()[name = string("op_2396_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2396_end_mask_0 = const()[name = string("op_2396_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_2396_cast_fp16 = slice_by_index(begin = var_2396_begin_0, end = var_2396_end_0, end_mask = var_2396_end_mask_0, x = cache_37_cast_fp16)[name = string("op_2396_cast_fp16")];
+            bool var_2402_interleave_0 = const()[name = string("op_2402_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_2402_cast_fp16 = concat(axis = var_69, interleave = var_2402_interleave_0, values = (var_2396_cast_fp16, key_19_cast_fp16))[name = string("op_2402_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195960832))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196747328))))[name = string("encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196747520)))];
+            tensor<fp16, [1, 14, 1024]> linear_84_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized, x = key_19_cast_fp16)[name = string("linear_84_cast_fp16")];
+            tensor<int32, [4]> var_2407 = const()[name = string("op_2407"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_55_cast_fp16 = reshape(shape = var_2407, x = linear_84_cast_fp16)[name = string("q_55_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196749632))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197536128))))[name = string("encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197536320)))];
+            tensor<fp16, [1, 56, 1024]> linear_85_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_85_cast_fp16")];
+            tensor<int32, [4]> var_2412 = const()[name = string("op_2412"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_37_cast_fp16 = reshape(shape = var_2412, x = linear_85_cast_fp16)[name = string("k_37_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197538432))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198324928))))[name = string("encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198325120)))];
+            tensor<fp16, [1, 56, 1024]> linear_86_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_86_cast_fp16")];
+            tensor<int32, [4]> var_2417 = const()[name = string("op_2417"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_19_cast_fp16 = reshape(shape = var_2417, x = linear_86_cast_fp16)[name = string("v_19_cast_fp16")];
+            tensor<int32, [4]> value_27_perm_0 = const()[name = string("value_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198327232)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2430_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = string("op_2430_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198329344)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2432_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = string("op_2432_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_19_perm_0 = const()[name = string("q_with_bias_v_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_241_transpose_x_0 = const()[name = string("x_241_transpose_x_0"), val = bool(false)];
+            bool x_241_transpose_y_0 = const()[name = string("x_241_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_2434_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198331456))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198445184))))[name = string("op_2434_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2432_cast_fp16)[name = string("transpose_281")];
+            tensor<fp16, [1, 8, 14, 111]> x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = op_2434_to_fp16_quantized)[name = string("x_241_cast_fp16")];
+            tensor<int32, [8]> x_243_pad_0 = const()[name = string("x_243_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_243_mode_0 = const()[name = string("x_243_mode_0"), val = string("constant")];
+            fp16 const_196_to_fp16 = const()[name = string("const_196_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_243_cast_fp16 = pad(constant_val = const_196_to_fp16, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241_cast_fp16)[name = string("x_243_cast_fp16")];
+            tensor<int32, [4]> var_2442 = const()[name = string("op_2442"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_245_cast_fp16 = reshape(shape = var_2442, x = x_243_cast_fp16)[name = string("x_245_cast_fp16")];
+            tensor<int32, [4]> var_2446_begin_0 = const()[name = string("op_2446_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2446_end_0 = const()[name = string("op_2446_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_2446_end_mask_0 = const()[name = string("op_2446_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_2446_cast_fp16 = slice_by_index(begin = var_2446_begin_0, end = var_2446_end_0, end_mask = var_2446_end_mask_0, x = x_245_cast_fp16)[name = string("op_2446_cast_fp16")];
+            tensor<int32, [4]> var_2447 = const()[name = string("op_2447"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_37_cast_fp16 = reshape(shape = var_2447, x = var_2446_cast_fp16)[name = string("matrix_bd_37_cast_fp16")];
+            bool matrix_ac_19_transpose_x_0 = const()[name = string("matrix_ac_19_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_19_transpose_y_0 = const()[name = string("matrix_ac_19_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_115 = transpose(perm = transpose_115_perm_0, x = k_37_cast_fp16)[name = string("transpose_279")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_114 = transpose(perm = transpose_114_perm_0, x = var_2430_cast_fp16)[name = string("transpose_280")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_19_cast_fp16 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_114, y = transpose_115)[name = string("matrix_ac_19_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_39_begin_0 = const()[name = string("matrix_bd_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_39_end_0 = const()[name = string("matrix_bd_39_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_39_end_mask_0 = const()[name = string("matrix_bd_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_39_cast_fp16 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37_cast_fp16)[name = string("matrix_bd_39_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2456_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = string("op_2456_cast_fp16")];
+            fp16 _inversed_scores_37_y_0_to_fp16 = const()[name = string("_inversed_scores_37_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_37_cast_fp16 = mul(x = var_2456_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = string("_inversed_scores_37_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_39_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_11)[name = string("scores_39_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2462_cast_fp16 = softmax(axis = var_60, x = scores_39_cast_fp16)[name = string("op_2462_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_509_cast_fp16 = select(a = var_45_to_fp16, b = var_2462_cast_fp16, cond = mask_11)[name = string("input_509_cast_fp16")];
+            bool x_247_transpose_x_0 = const()[name = string("x_247_transpose_x_0"), val = bool(false)];
+            bool x_247_transpose_y_0 = const()[name = string("x_247_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_19_cast_fp16)[name = string("transpose_278")];
+            tensor<fp16, [1, 8, 14, 128]> x_247_cast_fp16 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509_cast_fp16, y = value_27_cast_fp16)[name = string("x_247_cast_fp16")];
+            tensor<int32, [4]> var_2466_perm_0 = const()[name = string("op_2466_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2467 = const()[name = string("op_2467"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_2466_cast_fp16 = transpose(perm = var_2466_perm_0, x = x_247_cast_fp16)[name = string("transpose_277")];
+            tensor<fp16, [1, 14, 1024]> input_511_cast_fp16 = reshape(shape = var_2467, x = var_2466_cast_fp16)[name = string("input_511_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198445504))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199232000))))[name = string("encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199232192)))];
+            tensor<fp16, [1, 14, 1024]> linear_88_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized, x = input_511_cast_fp16)[name = string("linear_88_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = string("input_515_cast_fp16")];
+            tensor<int32, [1]> x_251_axes_0 = const()[name = string("x_251_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199234304)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199236416)))];
+            tensor<fp16, [1, 14, 1024]> x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = string("x_251_cast_fp16")];
+            tensor<int32, [3]> input_517_perm_0 = const()[name = string("input_517_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_519_pad_type_0 = const()[name = string("input_519_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_519_strides_0 = const()[name = string("input_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_519_pad_0 = const()[name = string("input_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_519_dilations_0 = const()[name = string("input_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_519_groups_0 = const()[name = string("input_519_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199238528))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201335744))))[name = string("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = string("transpose_276")];
+            tensor<fp16, [1, 2048, 14]> input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_517_cast_fp16)[name = string("input_519_cast_fp16")];
+            int32 x_253_split_num_splits_0 = const()[name = string("x_253_split_num_splits_0"), val = int32(2)];
+            int32 x_253_split_axis_0 = const()[name = string("x_253_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_253_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_253_split_cast_fp16_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519_cast_fp16)[name = string("x_253_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = string("x_253_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = string("x_253_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_521_cast_fp16 = select(a = var_45_to_fp16, b = x_253_cast_fp16, cond = var_576)[name = string("input_521_cast_fp16")];
+            bool new_x_39_interleave_0 = const()[name = string("new_x_39_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_39_cast_fp16 = concat(axis = var_60, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = string("new_x_39_cast_fp16")];
+            tensor<int32, [3]> var_2506_begin_0 = const()[name = string("op_2506_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_2506_end_0 = const()[name = string("op_2506_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_2506_end_mask_0 = const()[name = string("op_2506_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2506_cast_fp16 = slice_by_index(begin = var_2506_begin_0, end = var_2506_end_0, end_mask = var_2506_end_mask_0, x = new_x_39_cast_fp16)[name = string("op_2506_cast_fp16")];
+            string x_255_pad_type_0 = const()[name = string("x_255_pad_type_0"), val = string("valid")];
+            int32 x_255_groups_0 = const()[name = string("x_255_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_255_strides_0 = const()[name = string("x_255_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_255_pad_0 = const()[name = string("x_255_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_255_dilations_0 = const()[name = string("x_255_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201339904))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201349184))))[name = string("encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_39_cast_fp16)[name = string("x_255_cast_fp16")];
+            tensor<int32, [3]> input_523_perm_0 = const()[name = string("input_523_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_257_axes_0 = const()[name = string("x_257_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201351296)))];
+            tensor<fp16, [1024]> encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201353408)))];
+            tensor<fp16, [1, 14, 1024]> input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = string("transpose_275")];
+            tensor<fp16, [1, 14, 1024]> x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = string("x_257_cast_fp16")];
+            tensor<int32, [3]> input_525_perm_0 = const()[name = string("input_525_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = string("transpose_274")];
+            tensor<fp16, [1, 1024, 14]> input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = string("input_527_cast_fp16")];
+            string x_259_pad_type_0 = const()[name = string("x_259_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_259_strides_0 = const()[name = string("x_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_259_pad_0 = const()[name = string("x_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_259_dilations_0 = const()[name = string("x_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_259_groups_0 = const()[name = string("x_259_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201355520))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202404160))))[name = string("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_527_cast_fp16)[name = string("x_259_cast_fp16")];
+            tensor<int32, [3]> input_529_perm_0 = const()[name = string("input_529_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = string("transpose_273")];
+            tensor<fp16, [1, 14, 1024]> input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = string("input_531_cast_fp16")];
+            tensor<int32, [1]> input_533_axes_0 = const()[name = string("input_533_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202406272)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202408384)))];
+            tensor<fp16, [1, 14, 1024]> input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = string("input_533_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202410496))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205556288))))[name = string("encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_9_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205556480)))];
+            tensor<fp16, [1, 14, 4096]> linear_89_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized, x = input_533_cast_fp16)[name = string("linear_89_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = string("input_537_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205564736))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208710528))))[name = string("encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208710720)))];
+            tensor<fp16, [1, 14, 1024]> linear_90_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized, x = input_537_cast_fp16)[name = string("linear_90_cast_fp16")];
+            fp16 var_2549_to_fp16 = const()[name = string("op_2549_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_2550_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2549_to_fp16)[name = string("op_2550_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2550_cast_fp16)[name = string("input_543_cast_fp16")];
+            tensor<int32, [1]> input_545_axes_0 = const()[name = string("input_545_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208712832)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208714944)))];
+            tensor<fp16, [1, 14, 1024]> input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = string("input_545_cast_fp16")];
+            tensor<int32, [4]> cache_41_begin_0 = const()[name = string("cache_41_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_41_end_0 = const()[name = string("cache_41_end_0"), val = tensor<int32, [4]>([11, 1, 42, 1024])];
+            tensor<bool, [4]> cache_41_end_mask_0 = const()[name = string("cache_41_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_41_squeeze_mask_0 = const()[name = string("cache_41_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_41_cast_fp16 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_41_cast_fp16")];
+            tensor<int32, [4]> cache_43_begin_0 = const()[name = string("cache_43_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_43_end_0 = const()[name = string("cache_43_end_0"), val = tensor<int32, [4]>([11, 1, 1024, 8])];
+            tensor<bool, [4]> cache_43_end_mask_0 = const()[name = string("cache_43_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_43_squeeze_mask_0 = const()[name = string("cache_43_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_43_cast_fp16 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_43_cast_fp16")];
+            tensor<int32, [1]> input_547_axes_0 = const()[name = string("input_547_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208717056)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208719168)))];
+            tensor<fp16, [1, 14, 1024]> input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = string("input_547_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208721280))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211867072))))[name = string("encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_10_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211867264)))];
+            tensor<fp16, [1, 14, 4096]> linear_91_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized, x = input_547_cast_fp16)[name = string("linear_91_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = string("input_551_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211875520))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215021312))))[name = string("encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215021504)))];
+            tensor<fp16, [1, 14, 1024]> linear_92_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized, x = input_551_cast_fp16)[name = string("linear_92_cast_fp16")];
+            fp16 var_2586_to_fp16 = const()[name = string("op_2586_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_2587_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2586_to_fp16)[name = string("op_2587_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2587_cast_fp16)[name = string("input_557_cast_fp16")];
+            tensor<int32, [1]> key_21_axes_0 = const()[name = string("key_21_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215023616)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215025728)))];
+            tensor<fp16, [1, 14, 1024]> key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = string("key_21_cast_fp16")];
+            bool input_559_interleave_0 = const()[name = string("input_559_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_559_cast_fp16 = concat(axis = var_69, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = string("input_559_cast_fp16")];
+            tensor<int32, [3]> var_2609_begin_0 = const()[name = string("op_2609_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_2609_end_0 = const()[name = string("op_2609_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2609_end_mask_0 = const()[name = string("op_2609_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_2609_cast_fp16 = slice_by_index(begin = var_2609_begin_0, end = var_2609_end_0, end_mask = var_2609_end_mask_0, x = cache_41_cast_fp16)[name = string("op_2609_cast_fp16")];
+            bool var_2615_interleave_0 = const()[name = string("op_2615_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_2615_cast_fp16 = concat(axis = var_69, interleave = var_2615_interleave_0, values = (var_2609_cast_fp16, key_21_cast_fp16))[name = string("op_2615_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215027840))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215814336))))[name = string("encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215814528)))];
+            tensor<fp16, [1, 14, 1024]> linear_93_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized, x = key_21_cast_fp16)[name = string("linear_93_cast_fp16")];
+            tensor<int32, [4]> var_2620 = const()[name = string("op_2620"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_61_cast_fp16 = reshape(shape = var_2620, x = linear_93_cast_fp16)[name = string("q_61_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215816640))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216603136))))[name = string("encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216603328)))];
+            tensor<fp16, [1, 56, 1024]> linear_94_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_94_cast_fp16")];
+            tensor<int32, [4]> var_2625 = const()[name = string("op_2625"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_41_cast_fp16 = reshape(shape = var_2625, x = linear_94_cast_fp16)[name = string("k_41_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216605440))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217391936))))[name = string("encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217392128)))];
+            tensor<fp16, [1, 56, 1024]> linear_95_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_95_cast_fp16")];
+            tensor<int32, [4]> var_2630 = const()[name = string("op_2630"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_21_cast_fp16 = reshape(shape = var_2630, x = linear_95_cast_fp16)[name = string("v_21_cast_fp16")];
+            tensor<int32, [4]> value_29_perm_0 = const()[name = string("value_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217394240)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2643_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = string("op_2643_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217396352)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2645_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = string("op_2645_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_21_perm_0 = const()[name = string("q_with_bias_v_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_267_transpose_x_0 = const()[name = string("x_267_transpose_x_0"), val = bool(false)];
+            bool x_267_transpose_y_0 = const()[name = string("x_267_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_2647_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217398464))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217512192))))[name = string("op_2647_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2645_cast_fp16)[name = string("transpose_272")];
+            tensor<fp16, [1, 8, 14, 111]> x_267_cast_fp16 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = op_2647_to_fp16_quantized)[name = string("x_267_cast_fp16")];
+            tensor<int32, [8]> x_269_pad_0 = const()[name = string("x_269_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_269_mode_0 = const()[name = string("x_269_mode_0"), val = string("constant")];
+            fp16 const_209_to_fp16 = const()[name = string("const_209_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_269_cast_fp16 = pad(constant_val = const_209_to_fp16, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267_cast_fp16)[name = string("x_269_cast_fp16")];
+            tensor<int32, [4]> var_2655 = const()[name = string("op_2655"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_271_cast_fp16 = reshape(shape = var_2655, x = x_269_cast_fp16)[name = string("x_271_cast_fp16")];
+            tensor<int32, [4]> var_2659_begin_0 = const()[name = string("op_2659_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2659_end_0 = const()[name = string("op_2659_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_2659_end_mask_0 = const()[name = string("op_2659_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_2659_cast_fp16 = slice_by_index(begin = var_2659_begin_0, end = var_2659_end_0, end_mask = var_2659_end_mask_0, x = x_271_cast_fp16)[name = string("op_2659_cast_fp16")];
+            tensor<int32, [4]> var_2660 = const()[name = string("op_2660"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_41_cast_fp16 = reshape(shape = var_2660, x = var_2659_cast_fp16)[name = string("matrix_bd_41_cast_fp16")];
+            bool matrix_ac_21_transpose_x_0 = const()[name = string("matrix_ac_21_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_21_transpose_y_0 = const()[name = string("matrix_ac_21_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_117 = transpose(perm = transpose_117_perm_0, x = k_41_cast_fp16)[name = string("transpose_270")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_116 = transpose(perm = transpose_116_perm_0, x = var_2643_cast_fp16)[name = string("transpose_271")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_21_cast_fp16 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_116, y = transpose_117)[name = string("matrix_ac_21_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_43_begin_0 = const()[name = string("matrix_bd_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_43_end_0 = const()[name = string("matrix_bd_43_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_43_end_mask_0 = const()[name = string("matrix_bd_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_43_cast_fp16 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41_cast_fp16)[name = string("matrix_bd_43_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2669_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = string("op_2669_cast_fp16")];
+            fp16 _inversed_scores_41_y_0_to_fp16 = const()[name = string("_inversed_scores_41_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_41_cast_fp16 = mul(x = var_2669_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = string("_inversed_scores_41_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_43_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_11)[name = string("scores_43_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2675_cast_fp16 = softmax(axis = var_60, x = scores_43_cast_fp16)[name = string("op_2675_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_561_cast_fp16 = select(a = var_45_to_fp16, b = var_2675_cast_fp16, cond = mask_11)[name = string("input_561_cast_fp16")];
+            bool x_273_transpose_x_0 = const()[name = string("x_273_transpose_x_0"), val = bool(false)];
+            bool x_273_transpose_y_0 = const()[name = string("x_273_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_21_cast_fp16)[name = string("transpose_269")];
+            tensor<fp16, [1, 8, 14, 128]> x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561_cast_fp16, y = value_29_cast_fp16)[name = string("x_273_cast_fp16")];
+            tensor<int32, [4]> var_2679_perm_0 = const()[name = string("op_2679_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2680 = const()[name = string("op_2680"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_2679_cast_fp16 = transpose(perm = var_2679_perm_0, x = x_273_cast_fp16)[name = string("transpose_268")];
+            tensor<fp16, [1, 14, 1024]> input_563_cast_fp16 = reshape(shape = var_2680, x = var_2679_cast_fp16)[name = string("input_563_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217512512))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218299008))))[name = string("encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218299200)))];
+            tensor<fp16, [1, 14, 1024]> linear_97_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized, x = input_563_cast_fp16)[name = string("linear_97_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = string("input_567_cast_fp16")];
+            tensor<int32, [1]> x_277_axes_0 = const()[name = string("x_277_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218301312)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218303424)))];
+            tensor<fp16, [1, 14, 1024]> x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = string("x_277_cast_fp16")];
+            tensor<int32, [3]> input_569_perm_0 = const()[name = string("input_569_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_571_pad_type_0 = const()[name = string("input_571_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_571_strides_0 = const()[name = string("input_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_571_pad_0 = const()[name = string("input_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_571_dilations_0 = const()[name = string("input_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_571_groups_0 = const()[name = string("input_571_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218305536))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220402752))))[name = string("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = string("transpose_267")];
+            tensor<fp16, [1, 2048, 14]> input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_569_cast_fp16)[name = string("input_571_cast_fp16")];
+            int32 x_279_split_num_splits_0 = const()[name = string("x_279_split_num_splits_0"), val = int32(2)];
+            int32 x_279_split_axis_0 = const()[name = string("x_279_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_279_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_279_split_cast_fp16_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571_cast_fp16)[name = string("x_279_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = string("x_279_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = string("x_279_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_573_cast_fp16 = select(a = var_45_to_fp16, b = x_279_cast_fp16, cond = var_576)[name = string("input_573_cast_fp16")];
+            bool new_x_43_interleave_0 = const()[name = string("new_x_43_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_43_cast_fp16 = concat(axis = var_60, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = string("new_x_43_cast_fp16")];
+            tensor<int32, [3]> var_2719_begin_0 = const()[name = string("op_2719_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_2719_end_0 = const()[name = string("op_2719_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_2719_end_mask_0 = const()[name = string("op_2719_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2719_cast_fp16 = slice_by_index(begin = var_2719_begin_0, end = var_2719_end_0, end_mask = var_2719_end_mask_0, x = new_x_43_cast_fp16)[name = string("op_2719_cast_fp16")];
+            string x_281_pad_type_0 = const()[name = string("x_281_pad_type_0"), val = string("valid")];
+            int32 x_281_groups_0 = const()[name = string("x_281_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_281_strides_0 = const()[name = string("x_281_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_281_pad_0 = const()[name = string("x_281_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_281_dilations_0 = const()[name = string("x_281_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220406912))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220416192))))[name = string("encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_281_cast_fp16 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_43_cast_fp16)[name = string("x_281_cast_fp16")];
+            tensor<int32, [3]> input_575_perm_0 = const()[name = string("input_575_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_283_axes_0 = const()[name = string("x_283_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220418304)))];
+            tensor<fp16, [1024]> encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220420416)))];
+            tensor<fp16, [1, 14, 1024]> input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = string("transpose_266")];
+            tensor<fp16, [1, 14, 1024]> x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = string("x_283_cast_fp16")];
+            tensor<int32, [3]> input_577_perm_0 = const()[name = string("input_577_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = string("transpose_265")];
+            tensor<fp16, [1, 1024, 14]> input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = string("input_579_cast_fp16")];
+            string x_285_pad_type_0 = const()[name = string("x_285_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_285_strides_0 = const()[name = string("x_285_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_285_pad_0 = const()[name = string("x_285_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_285_dilations_0 = const()[name = string("x_285_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_285_groups_0 = const()[name = string("x_285_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220422528))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221471168))))[name = string("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_579_cast_fp16)[name = string("x_285_cast_fp16")];
+            tensor<int32, [3]> input_581_perm_0 = const()[name = string("input_581_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = string("transpose_264")];
+            tensor<fp16, [1, 14, 1024]> input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = string("input_583_cast_fp16")];
+            tensor<int32, [1]> input_585_axes_0 = const()[name = string("input_585_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221473280)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221475392)))];
+            tensor<fp16, [1, 14, 1024]> input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = string("input_585_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221477504))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224623296))))[name = string("encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_10_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224623488)))];
+            tensor<fp16, [1, 14, 4096]> linear_98_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized, x = input_585_cast_fp16)[name = string("linear_98_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = string("input_589_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224631744))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227777536))))[name = string("encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227777728)))];
+            tensor<fp16, [1, 14, 1024]> linear_99_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized, x = input_589_cast_fp16)[name = string("linear_99_cast_fp16")];
+            fp16 var_2762_to_fp16 = const()[name = string("op_2762_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_2763_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2762_to_fp16)[name = string("op_2763_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2763_cast_fp16)[name = string("input_595_cast_fp16")];
+            tensor<int32, [1]> input_597_axes_0 = const()[name = string("input_597_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227779840)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227781952)))];
+            tensor<fp16, [1, 14, 1024]> input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = string("input_597_cast_fp16")];
+            tensor<int32, [4]> cache_45_begin_0 = const()[name = string("cache_45_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_45_end_0 = const()[name = string("cache_45_end_0"), val = tensor<int32, [4]>([12, 1, 42, 1024])];
+            tensor<bool, [4]> cache_45_end_mask_0 = const()[name = string("cache_45_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_45_squeeze_mask_0 = const()[name = string("cache_45_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_45_cast_fp16 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_45_cast_fp16")];
+            tensor<int32, [4]> cache_47_begin_0 = const()[name = string("cache_47_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_47_end_0 = const()[name = string("cache_47_end_0"), val = tensor<int32, [4]>([12, 1, 1024, 8])];
+            tensor<bool, [4]> cache_47_end_mask_0 = const()[name = string("cache_47_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_47_squeeze_mask_0 = const()[name = string("cache_47_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_47_cast_fp16 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_47_cast_fp16")];
+            tensor<int32, [1]> input_599_axes_0 = const()[name = string("input_599_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227784064)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227786176)))];
+            tensor<fp16, [1, 14, 1024]> input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = string("input_599_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227788288))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230934080))))[name = string("encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_11_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230934272)))];
+            tensor<fp16, [1, 14, 4096]> linear_100_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized, x = input_599_cast_fp16)[name = string("linear_100_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = string("input_603_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230942528))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234088320))))[name = string("encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234088512)))];
+            tensor<fp16, [1, 14, 1024]> linear_101_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized, x = input_603_cast_fp16)[name = string("linear_101_cast_fp16")];
+            fp16 var_2799_to_fp16 = const()[name = string("op_2799_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_2800_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2799_to_fp16)[name = string("op_2800_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2800_cast_fp16)[name = string("input_609_cast_fp16")];
+            tensor<int32, [1]> key_23_axes_0 = const()[name = string("key_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234090624)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234092736)))];
+            tensor<fp16, [1, 14, 1024]> key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = string("key_23_cast_fp16")];
+            bool input_611_interleave_0 = const()[name = string("input_611_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_611_cast_fp16 = concat(axis = var_69, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = string("input_611_cast_fp16")];
+            tensor<int32, [3]> var_2822_begin_0 = const()[name = string("op_2822_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_2822_end_0 = const()[name = string("op_2822_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2822_end_mask_0 = const()[name = string("op_2822_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_2822_cast_fp16 = slice_by_index(begin = var_2822_begin_0, end = var_2822_end_0, end_mask = var_2822_end_mask_0, x = cache_45_cast_fp16)[name = string("op_2822_cast_fp16")];
+            bool var_2828_interleave_0 = const()[name = string("op_2828_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_2828_cast_fp16 = concat(axis = var_69, interleave = var_2828_interleave_0, values = (var_2822_cast_fp16, key_23_cast_fp16))[name = string("op_2828_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234094848))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234881344))))[name = string("encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234881536)))];
+            tensor<fp16, [1, 14, 1024]> linear_102_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized, x = key_23_cast_fp16)[name = string("linear_102_cast_fp16")];
+            tensor<int32, [4]> var_2833 = const()[name = string("op_2833"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_67_cast_fp16 = reshape(shape = var_2833, x = linear_102_cast_fp16)[name = string("q_67_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234883648))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235670144))))[name = string("encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235670336)))];
+            tensor<fp16, [1, 56, 1024]> linear_103_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_103_cast_fp16")];
+            tensor<int32, [4]> var_2838 = const()[name = string("op_2838"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_45_cast_fp16 = reshape(shape = var_2838, x = linear_103_cast_fp16)[name = string("k_45_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235672448))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236458944))))[name = string("encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236459136)))];
+            tensor<fp16, [1, 56, 1024]> linear_104_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_104_cast_fp16")];
+            tensor<int32, [4]> var_2843 = const()[name = string("op_2843"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_23_cast_fp16 = reshape(shape = var_2843, x = linear_104_cast_fp16)[name = string("v_23_cast_fp16")];
+            tensor<int32, [4]> value_31_perm_0 = const()[name = string("value_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236461248)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2856_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = string("op_2856_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236463360)))];
+            tensor<fp16, [1, 14, 8, 128]> var_2858_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = string("op_2858_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_23_perm_0 = const()[name = string("q_with_bias_v_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_293_transpose_x_0 = const()[name = string("x_293_transpose_x_0"), val = bool(false)];
+            bool x_293_transpose_y_0 = const()[name = string("x_293_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_2860_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236465472))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236579200))))[name = string("op_2860_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2858_cast_fp16)[name = string("transpose_263")];
+            tensor<fp16, [1, 8, 14, 111]> x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = op_2860_to_fp16_quantized)[name = string("x_293_cast_fp16")];
+            tensor<int32, [8]> x_295_pad_0 = const()[name = string("x_295_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_295_mode_0 = const()[name = string("x_295_mode_0"), val = string("constant")];
+            fp16 const_222_to_fp16 = const()[name = string("const_222_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_295_cast_fp16 = pad(constant_val = const_222_to_fp16, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293_cast_fp16)[name = string("x_295_cast_fp16")];
+            tensor<int32, [4]> var_2868 = const()[name = string("op_2868"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_297_cast_fp16 = reshape(shape = var_2868, x = x_295_cast_fp16)[name = string("x_297_cast_fp16")];
+            tensor<int32, [4]> var_2872_begin_0 = const()[name = string("op_2872_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2872_end_0 = const()[name = string("op_2872_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_2872_end_mask_0 = const()[name = string("op_2872_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_2872_cast_fp16 = slice_by_index(begin = var_2872_begin_0, end = var_2872_end_0, end_mask = var_2872_end_mask_0, x = x_297_cast_fp16)[name = string("op_2872_cast_fp16")];
+            tensor<int32, [4]> var_2873 = const()[name = string("op_2873"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_45_cast_fp16 = reshape(shape = var_2873, x = var_2872_cast_fp16)[name = string("matrix_bd_45_cast_fp16")];
+            bool matrix_ac_23_transpose_x_0 = const()[name = string("matrix_ac_23_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_23_transpose_y_0 = const()[name = string("matrix_ac_23_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_119 = transpose(perm = transpose_119_perm_0, x = k_45_cast_fp16)[name = string("transpose_261")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_118 = transpose(perm = transpose_118_perm_0, x = var_2856_cast_fp16)[name = string("transpose_262")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_23_cast_fp16 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_118, y = transpose_119)[name = string("matrix_ac_23_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_47_begin_0 = const()[name = string("matrix_bd_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_47_end_0 = const()[name = string("matrix_bd_47_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_47_end_mask_0 = const()[name = string("matrix_bd_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_47_cast_fp16 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45_cast_fp16)[name = string("matrix_bd_47_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2882_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = string("op_2882_cast_fp16")];
+            fp16 _inversed_scores_45_y_0_to_fp16 = const()[name = string("_inversed_scores_45_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_45_cast_fp16 = mul(x = var_2882_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = string("_inversed_scores_45_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_47_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_11)[name = string("scores_47_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_2888_cast_fp16 = softmax(axis = var_60, x = scores_47_cast_fp16)[name = string("op_2888_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_613_cast_fp16 = select(a = var_45_to_fp16, b = var_2888_cast_fp16, cond = mask_11)[name = string("input_613_cast_fp16")];
+            bool x_299_transpose_x_0 = const()[name = string("x_299_transpose_x_0"), val = bool(false)];
+            bool x_299_transpose_y_0 = const()[name = string("x_299_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_23_cast_fp16)[name = string("transpose_260")];
+            tensor<fp16, [1, 8, 14, 128]> x_299_cast_fp16 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613_cast_fp16, y = value_31_cast_fp16)[name = string("x_299_cast_fp16")];
+            tensor<int32, [4]> var_2892_perm_0 = const()[name = string("op_2892_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2893 = const()[name = string("op_2893"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_2892_cast_fp16 = transpose(perm = var_2892_perm_0, x = x_299_cast_fp16)[name = string("transpose_259")];
+            tensor<fp16, [1, 14, 1024]> input_615_cast_fp16 = reshape(shape = var_2893, x = var_2892_cast_fp16)[name = string("input_615_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236579520))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237366016))))[name = string("encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237366208)))];
+            tensor<fp16, [1, 14, 1024]> linear_106_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized, x = input_615_cast_fp16)[name = string("linear_106_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = string("input_619_cast_fp16")];
+            tensor<int32, [1]> x_303_axes_0 = const()[name = string("x_303_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237368320)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237370432)))];
+            tensor<fp16, [1, 14, 1024]> x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = string("x_303_cast_fp16")];
+            tensor<int32, [3]> input_621_perm_0 = const()[name = string("input_621_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_623_pad_type_0 = const()[name = string("input_623_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_623_strides_0 = const()[name = string("input_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_623_pad_0 = const()[name = string("input_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_623_dilations_0 = const()[name = string("input_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_623_groups_0 = const()[name = string("input_623_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237372544))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239469760))))[name = string("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = string("transpose_258")];
+            tensor<fp16, [1, 2048, 14]> input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_621_cast_fp16)[name = string("input_623_cast_fp16")];
+            int32 x_305_split_num_splits_0 = const()[name = string("x_305_split_num_splits_0"), val = int32(2)];
+            int32 x_305_split_axis_0 = const()[name = string("x_305_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_305_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623_cast_fp16)[name = string("x_305_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = string("x_305_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = string("x_305_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_625_cast_fp16 = select(a = var_45_to_fp16, b = x_305_cast_fp16, cond = var_576)[name = string("input_625_cast_fp16")];
+            bool new_x_47_interleave_0 = const()[name = string("new_x_47_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_47_cast_fp16 = concat(axis = var_60, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = string("new_x_47_cast_fp16")];
+            tensor<int32, [3]> var_2932_begin_0 = const()[name = string("op_2932_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_2932_end_0 = const()[name = string("op_2932_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_2932_end_mask_0 = const()[name = string("op_2932_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2932_cast_fp16 = slice_by_index(begin = var_2932_begin_0, end = var_2932_end_0, end_mask = var_2932_end_mask_0, x = new_x_47_cast_fp16)[name = string("op_2932_cast_fp16")];
+            string x_307_pad_type_0 = const()[name = string("x_307_pad_type_0"), val = string("valid")];
+            int32 x_307_groups_0 = const()[name = string("x_307_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_307_strides_0 = const()[name = string("x_307_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_307_pad_0 = const()[name = string("x_307_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_307_dilations_0 = const()[name = string("x_307_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239473920))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239483200))))[name = string("encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_47_cast_fp16)[name = string("x_307_cast_fp16")];
+            tensor<int32, [3]> input_627_perm_0 = const()[name = string("input_627_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_309_axes_0 = const()[name = string("x_309_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239485312)))];
+            tensor<fp16, [1024]> encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239487424)))];
+            tensor<fp16, [1, 14, 1024]> input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = string("transpose_257")];
+            tensor<fp16, [1, 14, 1024]> x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = string("x_309_cast_fp16")];
+            tensor<int32, [3]> input_629_perm_0 = const()[name = string("input_629_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = string("transpose_256")];
+            tensor<fp16, [1, 1024, 14]> input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = string("input_631_cast_fp16")];
+            string x_311_pad_type_0 = const()[name = string("x_311_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_311_strides_0 = const()[name = string("x_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_311_pad_0 = const()[name = string("x_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_311_dilations_0 = const()[name = string("x_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_311_groups_0 = const()[name = string("x_311_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239489536))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240538176))))[name = string("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_311_cast_fp16 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_631_cast_fp16)[name = string("x_311_cast_fp16")];
+            tensor<int32, [3]> input_633_perm_0 = const()[name = string("input_633_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = string("transpose_255")];
+            tensor<fp16, [1, 14, 1024]> input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = string("input_635_cast_fp16")];
+            tensor<int32, [1]> input_637_axes_0 = const()[name = string("input_637_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240540288)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240542400)))];
+            tensor<fp16, [1, 14, 1024]> input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = string("input_637_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240544512))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243690304))))[name = string("encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_11_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243690496)))];
+            tensor<fp16, [1, 14, 4096]> linear_107_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized, x = input_637_cast_fp16)[name = string("linear_107_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = string("input_641_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243698752))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246844544))))[name = string("encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246844736)))];
+            tensor<fp16, [1, 14, 1024]> linear_108_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized, x = input_641_cast_fp16)[name = string("linear_108_cast_fp16")];
+            fp16 var_2975_to_fp16 = const()[name = string("op_2975_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_2976_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2975_to_fp16)[name = string("op_2976_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2976_cast_fp16)[name = string("input_647_cast_fp16")];
+            tensor<int32, [1]> input_649_axes_0 = const()[name = string("input_649_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246846848)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246848960)))];
+            tensor<fp16, [1, 14, 1024]> input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = string("input_649_cast_fp16")];
+            tensor<int32, [4]> cache_49_begin_0 = const()[name = string("cache_49_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_49_end_0 = const()[name = string("cache_49_end_0"), val = tensor<int32, [4]>([13, 1, 42, 1024])];
+            tensor<bool, [4]> cache_49_end_mask_0 = const()[name = string("cache_49_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_49_squeeze_mask_0 = const()[name = string("cache_49_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_49_cast_fp16 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_49_cast_fp16")];
+            tensor<int32, [4]> cache_51_begin_0 = const()[name = string("cache_51_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_51_end_0 = const()[name = string("cache_51_end_0"), val = tensor<int32, [4]>([13, 1, 1024, 8])];
+            tensor<bool, [4]> cache_51_end_mask_0 = const()[name = string("cache_51_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_51_squeeze_mask_0 = const()[name = string("cache_51_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_51_cast_fp16 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_51_cast_fp16")];
+            tensor<int32, [1]> input_651_axes_0 = const()[name = string("input_651_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246851072)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246853184)))];
+            tensor<fp16, [1, 14, 1024]> input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = string("input_651_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246855296))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250001088))))[name = string("encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_12_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250001280)))];
+            tensor<fp16, [1, 14, 4096]> linear_109_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized, x = input_651_cast_fp16)[name = string("linear_109_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("input_655_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250009536))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253155328))))[name = string("encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253155520)))];
+            tensor<fp16, [1, 14, 1024]> linear_110_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized, x = input_655_cast_fp16)[name = string("linear_110_cast_fp16")];
+            fp16 var_3012_to_fp16 = const()[name = string("op_3012_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_3013_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_3012_to_fp16)[name = string("op_3013_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_3013_cast_fp16)[name = string("input_661_cast_fp16")];
+            tensor<int32, [1]> key_25_axes_0 = const()[name = string("key_25_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253157632)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253159744)))];
+            tensor<fp16, [1, 14, 1024]> key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = string("key_25_cast_fp16")];
+            bool input_663_interleave_0 = const()[name = string("input_663_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_663_cast_fp16 = concat(axis = var_69, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = string("input_663_cast_fp16")];
+            tensor<int32, [3]> var_3035_begin_0 = const()[name = string("op_3035_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_3035_end_0 = const()[name = string("op_3035_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3035_end_mask_0 = const()[name = string("op_3035_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_3035_cast_fp16 = slice_by_index(begin = var_3035_begin_0, end = var_3035_end_0, end_mask = var_3035_end_mask_0, x = cache_49_cast_fp16)[name = string("op_3035_cast_fp16")];
+            bool var_3041_interleave_0 = const()[name = string("op_3041_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3041_cast_fp16 = concat(axis = var_69, interleave = var_3041_interleave_0, values = (var_3035_cast_fp16, key_25_cast_fp16))[name = string("op_3041_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253161856))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253948352))))[name = string("encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253948544)))];
+            tensor<fp16, [1, 14, 1024]> linear_111_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized, x = key_25_cast_fp16)[name = string("linear_111_cast_fp16")];
+            tensor<int32, [4]> var_3046 = const()[name = string("op_3046"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_73_cast_fp16 = reshape(shape = var_3046, x = linear_111_cast_fp16)[name = string("q_73_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253950656))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254737152))))[name = string("encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254737344)))];
+            tensor<fp16, [1, 56, 1024]> linear_112_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_112_cast_fp16")];
+            tensor<int32, [4]> var_3051 = const()[name = string("op_3051"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_49_cast_fp16 = reshape(shape = var_3051, x = linear_112_cast_fp16)[name = string("k_49_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254739456))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255525952))))[name = string("encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255526144)))];
+            tensor<fp16, [1, 56, 1024]> linear_113_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_113_cast_fp16")];
+            tensor<int32, [4]> var_3056 = const()[name = string("op_3056"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_25_cast_fp16 = reshape(shape = var_3056, x = linear_113_cast_fp16)[name = string("v_25_cast_fp16")];
+            tensor<int32, [4]> value_33_perm_0 = const()[name = string("value_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255528256)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3069_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = string("op_3069_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255530368)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3071_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = string("op_3071_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_25_perm_0 = const()[name = string("q_with_bias_v_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_319_transpose_x_0 = const()[name = string("x_319_transpose_x_0"), val = bool(false)];
+            bool x_319_transpose_y_0 = const()[name = string("x_319_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_3073_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255532480))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255646208))))[name = string("op_3073_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_3071_cast_fp16)[name = string("transpose_254")];
+            tensor<fp16, [1, 8, 14, 111]> x_319_cast_fp16 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = op_3073_to_fp16_quantized)[name = string("x_319_cast_fp16")];
+            tensor<int32, [8]> x_321_pad_0 = const()[name = string("x_321_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("constant")];
+            fp16 const_235_to_fp16 = const()[name = string("const_235_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_321_cast_fp16 = pad(constant_val = const_235_to_fp16, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319_cast_fp16)[name = string("x_321_cast_fp16")];
+            tensor<int32, [4]> var_3081 = const()[name = string("op_3081"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_323_cast_fp16 = reshape(shape = var_3081, x = x_321_cast_fp16)[name = string("x_323_cast_fp16")];
+            tensor<int32, [4]> var_3085_begin_0 = const()[name = string("op_3085_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3085_end_0 = const()[name = string("op_3085_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_3085_end_mask_0 = const()[name = string("op_3085_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_3085_cast_fp16 = slice_by_index(begin = var_3085_begin_0, end = var_3085_end_0, end_mask = var_3085_end_mask_0, x = x_323_cast_fp16)[name = string("op_3085_cast_fp16")];
+            tensor<int32, [4]> var_3086 = const()[name = string("op_3086"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_49_cast_fp16 = reshape(shape = var_3086, x = var_3085_cast_fp16)[name = string("matrix_bd_49_cast_fp16")];
+            bool matrix_ac_25_transpose_x_0 = const()[name = string("matrix_ac_25_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_25_transpose_y_0 = const()[name = string("matrix_ac_25_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_120_perm_0 = const()[name = string("transpose_120_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_121_perm_0 = const()[name = string("transpose_121_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_121 = transpose(perm = transpose_121_perm_0, x = k_49_cast_fp16)[name = string("transpose_252")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_120 = transpose(perm = transpose_120_perm_0, x = var_3069_cast_fp16)[name = string("transpose_253")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_25_cast_fp16 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_120, y = transpose_121)[name = string("matrix_ac_25_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_51_begin_0 = const()[name = string("matrix_bd_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_51_end_0 = const()[name = string("matrix_bd_51_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_51_end_mask_0 = const()[name = string("matrix_bd_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_51_cast_fp16 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49_cast_fp16)[name = string("matrix_bd_51_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3095_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = string("op_3095_cast_fp16")];
+            fp16 _inversed_scores_49_y_0_to_fp16 = const()[name = string("_inversed_scores_49_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_49_cast_fp16 = mul(x = var_3095_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = string("_inversed_scores_49_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_51_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_11)[name = string("scores_51_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3101_cast_fp16 = softmax(axis = var_60, x = scores_51_cast_fp16)[name = string("op_3101_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_665_cast_fp16 = select(a = var_45_to_fp16, b = var_3101_cast_fp16, cond = mask_11)[name = string("input_665_cast_fp16")];
+            bool x_325_transpose_x_0 = const()[name = string("x_325_transpose_x_0"), val = bool(false)];
+            bool x_325_transpose_y_0 = const()[name = string("x_325_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_25_cast_fp16)[name = string("transpose_251")];
+            tensor<fp16, [1, 8, 14, 128]> x_325_cast_fp16 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665_cast_fp16, y = value_33_cast_fp16)[name = string("x_325_cast_fp16")];
+            tensor<int32, [4]> var_3105_perm_0 = const()[name = string("op_3105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3106 = const()[name = string("op_3106"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_3105_cast_fp16 = transpose(perm = var_3105_perm_0, x = x_325_cast_fp16)[name = string("transpose_250")];
+            tensor<fp16, [1, 14, 1024]> input_667_cast_fp16 = reshape(shape = var_3106, x = var_3105_cast_fp16)[name = string("input_667_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255646528))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256433024))))[name = string("encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256433216)))];
+            tensor<fp16, [1, 14, 1024]> linear_115_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized, x = input_667_cast_fp16)[name = string("linear_115_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = string("input_671_cast_fp16")];
+            tensor<int32, [1]> x_329_axes_0 = const()[name = string("x_329_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256435328)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256437440)))];
+            tensor<fp16, [1, 14, 1024]> x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = string("x_329_cast_fp16")];
+            tensor<int32, [3]> input_673_perm_0 = const()[name = string("input_673_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_675_pad_type_0 = const()[name = string("input_675_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_675_strides_0 = const()[name = string("input_675_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_675_pad_0 = const()[name = string("input_675_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_675_dilations_0 = const()[name = string("input_675_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_675_groups_0 = const()[name = string("input_675_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256439552))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258536768))))[name = string("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = string("transpose_249")];
+            tensor<fp16, [1, 2048, 14]> input_675_cast_fp16 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_673_cast_fp16)[name = string("input_675_cast_fp16")];
+            int32 x_331_split_num_splits_0 = const()[name = string("x_331_split_num_splits_0"), val = int32(2)];
+            int32 x_331_split_axis_0 = const()[name = string("x_331_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_331_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_331_split_cast_fp16_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675_cast_fp16)[name = string("x_331_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = string("x_331_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = string("x_331_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_677_cast_fp16 = select(a = var_45_to_fp16, b = x_331_cast_fp16, cond = var_576)[name = string("input_677_cast_fp16")];
+            bool new_x_51_interleave_0 = const()[name = string("new_x_51_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_51_cast_fp16 = concat(axis = var_60, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = string("new_x_51_cast_fp16")];
+            tensor<int32, [3]> var_3145_begin_0 = const()[name = string("op_3145_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_3145_end_0 = const()[name = string("op_3145_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_3145_end_mask_0 = const()[name = string("op_3145_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3145_cast_fp16 = slice_by_index(begin = var_3145_begin_0, end = var_3145_end_0, end_mask = var_3145_end_mask_0, x = new_x_51_cast_fp16)[name = string("op_3145_cast_fp16")];
+            string x_333_pad_type_0 = const()[name = string("x_333_pad_type_0"), val = string("valid")];
+            int32 x_333_groups_0 = const()[name = string("x_333_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_333_strides_0 = const()[name = string("x_333_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_333_pad_0 = const()[name = string("x_333_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_333_dilations_0 = const()[name = string("x_333_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258540928))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258550208))))[name = string("encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_333_cast_fp16 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_51_cast_fp16)[name = string("x_333_cast_fp16")];
+            tensor<int32, [3]> input_679_perm_0 = const()[name = string("input_679_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_335_axes_0 = const()[name = string("x_335_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258552320)))];
+            tensor<fp16, [1024]> encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258554432)))];
+            tensor<fp16, [1, 14, 1024]> input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = string("transpose_248")];
+            tensor<fp16, [1, 14, 1024]> x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = string("x_335_cast_fp16")];
+            tensor<int32, [3]> input_681_perm_0 = const()[name = string("input_681_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = string("transpose_247")];
+            tensor<fp16, [1, 1024, 14]> input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = string("input_683_cast_fp16")];
+            string x_337_pad_type_0 = const()[name = string("x_337_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_337_strides_0 = const()[name = string("x_337_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_337_pad_0 = const()[name = string("x_337_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_337_dilations_0 = const()[name = string("x_337_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_337_groups_0 = const()[name = string("x_337_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258556544))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259605184))))[name = string("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_337_cast_fp16 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_683_cast_fp16)[name = string("x_337_cast_fp16")];
+            tensor<int32, [3]> input_685_perm_0 = const()[name = string("input_685_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = string("transpose_246")];
+            tensor<fp16, [1, 14, 1024]> input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = string("input_687_cast_fp16")];
+            tensor<int32, [1]> input_689_axes_0 = const()[name = string("input_689_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259607296)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259609408)))];
+            tensor<fp16, [1, 14, 1024]> input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = string("input_689_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259611520))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262757312))))[name = string("encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_12_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262757504)))];
+            tensor<fp16, [1, 14, 4096]> linear_116_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized, x = input_689_cast_fp16)[name = string("linear_116_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("input_693_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262765760))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265911552))))[name = string("encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265911744)))];
+            tensor<fp16, [1, 14, 1024]> linear_117_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized, x = input_693_cast_fp16)[name = string("linear_117_cast_fp16")];
+            fp16 var_3188_to_fp16 = const()[name = string("op_3188_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_3189_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_3188_to_fp16)[name = string("op_3189_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_3189_cast_fp16)[name = string("input_699_cast_fp16")];
+            tensor<int32, [1]> input_701_axes_0 = const()[name = string("input_701_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265913856)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265915968)))];
+            tensor<fp16, [1, 14, 1024]> input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = string("input_701_cast_fp16")];
+            tensor<int32, [4]> cache_53_begin_0 = const()[name = string("cache_53_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_53_end_0 = const()[name = string("cache_53_end_0"), val = tensor<int32, [4]>([14, 1, 42, 1024])];
+            tensor<bool, [4]> cache_53_end_mask_0 = const()[name = string("cache_53_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_53_squeeze_mask_0 = const()[name = string("cache_53_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_53_cast_fp16 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_53_cast_fp16")];
+            tensor<int32, [4]> cache_55_begin_0 = const()[name = string("cache_55_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_55_end_0 = const()[name = string("cache_55_end_0"), val = tensor<int32, [4]>([14, 1, 1024, 8])];
+            tensor<bool, [4]> cache_55_end_mask_0 = const()[name = string("cache_55_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_55_squeeze_mask_0 = const()[name = string("cache_55_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_55_cast_fp16 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_55_cast_fp16")];
+            tensor<int32, [1]> input_703_axes_0 = const()[name = string("input_703_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265918080)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265920192)))];
+            tensor<fp16, [1, 14, 1024]> input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = string("input_703_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265922304))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269068096))))[name = string("encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_13_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269068288)))];
+            tensor<fp16, [1, 14, 4096]> linear_118_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized, x = input_703_cast_fp16)[name = string("linear_118_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = string("input_707_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269076544))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272222336))))[name = string("encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272222528)))];
+            tensor<fp16, [1, 14, 1024]> linear_119_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized, x = input_707_cast_fp16)[name = string("linear_119_cast_fp16")];
+            fp16 var_3225_to_fp16 = const()[name = string("op_3225_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_3226_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_3225_to_fp16)[name = string("op_3226_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_3226_cast_fp16)[name = string("input_713_cast_fp16")];
+            tensor<int32, [1]> key_27_axes_0 = const()[name = string("key_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272224640)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272226752)))];
+            tensor<fp16, [1, 14, 1024]> key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = string("key_27_cast_fp16")];
+            bool input_715_interleave_0 = const()[name = string("input_715_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_715_cast_fp16 = concat(axis = var_69, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = string("input_715_cast_fp16")];
+            tensor<int32, [3]> var_3248_begin_0 = const()[name = string("op_3248_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_3248_end_0 = const()[name = string("op_3248_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3248_end_mask_0 = const()[name = string("op_3248_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_3248_cast_fp16 = slice_by_index(begin = var_3248_begin_0, end = var_3248_end_0, end_mask = var_3248_end_mask_0, x = cache_53_cast_fp16)[name = string("op_3248_cast_fp16")];
+            bool var_3254_interleave_0 = const()[name = string("op_3254_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3254_cast_fp16 = concat(axis = var_69, interleave = var_3254_interleave_0, values = (var_3248_cast_fp16, key_27_cast_fp16))[name = string("op_3254_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272228864))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273015360))))[name = string("encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273015552)))];
+            tensor<fp16, [1, 14, 1024]> linear_120_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized, x = key_27_cast_fp16)[name = string("linear_120_cast_fp16")];
+            tensor<int32, [4]> var_3259 = const()[name = string("op_3259"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_79_cast_fp16 = reshape(shape = var_3259, x = linear_120_cast_fp16)[name = string("q_79_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273017664))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273804160))))[name = string("encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273804352)))];
+            tensor<fp16, [1, 56, 1024]> linear_121_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_121_cast_fp16")];
+            tensor<int32, [4]> var_3264 = const()[name = string("op_3264"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_53_cast_fp16 = reshape(shape = var_3264, x = linear_121_cast_fp16)[name = string("k_53_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273806464))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274592960))))[name = string("encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274593152)))];
+            tensor<fp16, [1, 56, 1024]> linear_122_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_122_cast_fp16")];
+            tensor<int32, [4]> var_3269 = const()[name = string("op_3269"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_27_cast_fp16 = reshape(shape = var_3269, x = linear_122_cast_fp16)[name = string("v_27_cast_fp16")];
+            tensor<int32, [4]> value_35_perm_0 = const()[name = string("value_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274595264)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3282_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = string("op_3282_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274597376)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3284_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = string("op_3284_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_27_perm_0 = const()[name = string("q_with_bias_v_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_345_transpose_x_0 = const()[name = string("x_345_transpose_x_0"), val = bool(false)];
+            bool x_345_transpose_y_0 = const()[name = string("x_345_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_3286_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274599488))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274713216))))[name = string("op_3286_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3284_cast_fp16)[name = string("transpose_245")];
+            tensor<fp16, [1, 8, 14, 111]> x_345_cast_fp16 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = op_3286_to_fp16_quantized)[name = string("x_345_cast_fp16")];
+            tensor<int32, [8]> x_347_pad_0 = const()[name = string("x_347_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_347_mode_0 = const()[name = string("x_347_mode_0"), val = string("constant")];
+            fp16 const_248_to_fp16 = const()[name = string("const_248_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_347_cast_fp16 = pad(constant_val = const_248_to_fp16, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345_cast_fp16)[name = string("x_347_cast_fp16")];
+            tensor<int32, [4]> var_3294 = const()[name = string("op_3294"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_349_cast_fp16 = reshape(shape = var_3294, x = x_347_cast_fp16)[name = string("x_349_cast_fp16")];
+            tensor<int32, [4]> var_3298_begin_0 = const()[name = string("op_3298_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3298_end_0 = const()[name = string("op_3298_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_3298_end_mask_0 = const()[name = string("op_3298_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_3298_cast_fp16 = slice_by_index(begin = var_3298_begin_0, end = var_3298_end_0, end_mask = var_3298_end_mask_0, x = x_349_cast_fp16)[name = string("op_3298_cast_fp16")];
+            tensor<int32, [4]> var_3299 = const()[name = string("op_3299"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_53_cast_fp16 = reshape(shape = var_3299, x = var_3298_cast_fp16)[name = string("matrix_bd_53_cast_fp16")];
+            bool matrix_ac_27_transpose_x_0 = const()[name = string("matrix_ac_27_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_27_transpose_y_0 = const()[name = string("matrix_ac_27_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_122_perm_0 = const()[name = string("transpose_122_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_123_perm_0 = const()[name = string("transpose_123_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_123 = transpose(perm = transpose_123_perm_0, x = k_53_cast_fp16)[name = string("transpose_243")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_122 = transpose(perm = transpose_122_perm_0, x = var_3282_cast_fp16)[name = string("transpose_244")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_27_cast_fp16 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_122, y = transpose_123)[name = string("matrix_ac_27_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_55_begin_0 = const()[name = string("matrix_bd_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_55_end_0 = const()[name = string("matrix_bd_55_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_55_end_mask_0 = const()[name = string("matrix_bd_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_55_cast_fp16 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53_cast_fp16)[name = string("matrix_bd_55_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3308_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = string("op_3308_cast_fp16")];
+            fp16 _inversed_scores_53_y_0_to_fp16 = const()[name = string("_inversed_scores_53_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_53_cast_fp16 = mul(x = var_3308_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = string("_inversed_scores_53_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_55_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_11)[name = string("scores_55_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3314_cast_fp16 = softmax(axis = var_60, x = scores_55_cast_fp16)[name = string("op_3314_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_717_cast_fp16 = select(a = var_45_to_fp16, b = var_3314_cast_fp16, cond = mask_11)[name = string("input_717_cast_fp16")];
+            bool x_351_transpose_x_0 = const()[name = string("x_351_transpose_x_0"), val = bool(false)];
+            bool x_351_transpose_y_0 = const()[name = string("x_351_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = v_27_cast_fp16)[name = string("transpose_242")];
+            tensor<fp16, [1, 8, 14, 128]> x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717_cast_fp16, y = value_35_cast_fp16)[name = string("x_351_cast_fp16")];
+            tensor<int32, [4]> var_3318_perm_0 = const()[name = string("op_3318_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3319 = const()[name = string("op_3319"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_3318_cast_fp16 = transpose(perm = var_3318_perm_0, x = x_351_cast_fp16)[name = string("transpose_241")];
+            tensor<fp16, [1, 14, 1024]> input_719_cast_fp16 = reshape(shape = var_3319, x = var_3318_cast_fp16)[name = string("input_719_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274713536))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275500032))))[name = string("encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275500224)))];
+            tensor<fp16, [1, 14, 1024]> linear_124_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("linear_124_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = string("input_723_cast_fp16")];
+            tensor<int32, [1]> x_355_axes_0 = const()[name = string("x_355_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275502336)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275504448)))];
+            tensor<fp16, [1, 14, 1024]> x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = string("x_355_cast_fp16")];
+            tensor<int32, [3]> input_725_perm_0 = const()[name = string("input_725_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_727_pad_type_0 = const()[name = string("input_727_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_727_strides_0 = const()[name = string("input_727_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_727_pad_0 = const()[name = string("input_727_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_727_dilations_0 = const()[name = string("input_727_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_727_groups_0 = const()[name = string("input_727_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275506560))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277603776))))[name = string("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = string("transpose_240")];
+            tensor<fp16, [1, 2048, 14]> input_727_cast_fp16 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_725_cast_fp16)[name = string("input_727_cast_fp16")];
+            int32 x_357_split_num_splits_0 = const()[name = string("x_357_split_num_splits_0"), val = int32(2)];
+            int32 x_357_split_axis_0 = const()[name = string("x_357_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_357_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727_cast_fp16)[name = string("x_357_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = string("x_357_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = string("x_357_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_729_cast_fp16 = select(a = var_45_to_fp16, b = x_357_cast_fp16, cond = var_576)[name = string("input_729_cast_fp16")];
+            bool new_x_55_interleave_0 = const()[name = string("new_x_55_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_55_cast_fp16 = concat(axis = var_60, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = string("new_x_55_cast_fp16")];
+            tensor<int32, [3]> var_3358_begin_0 = const()[name = string("op_3358_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_3358_end_0 = const()[name = string("op_3358_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_3358_end_mask_0 = const()[name = string("op_3358_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3358_cast_fp16 = slice_by_index(begin = var_3358_begin_0, end = var_3358_end_0, end_mask = var_3358_end_mask_0, x = new_x_55_cast_fp16)[name = string("op_3358_cast_fp16")];
+            string x_359_pad_type_0 = const()[name = string("x_359_pad_type_0"), val = string("valid")];
+            int32 x_359_groups_0 = const()[name = string("x_359_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_359_strides_0 = const()[name = string("x_359_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_359_pad_0 = const()[name = string("x_359_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_359_dilations_0 = const()[name = string("x_359_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277607936))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277617216))))[name = string("encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_55_cast_fp16)[name = string("x_359_cast_fp16")];
+            tensor<int32, [3]> input_731_perm_0 = const()[name = string("input_731_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_361_axes_0 = const()[name = string("x_361_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277619328)))];
+            tensor<fp16, [1024]> encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277621440)))];
+            tensor<fp16, [1, 14, 1024]> input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = string("transpose_239")];
+            tensor<fp16, [1, 14, 1024]> x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = string("x_361_cast_fp16")];
+            tensor<int32, [3]> input_733_perm_0 = const()[name = string("input_733_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = string("transpose_238")];
+            tensor<fp16, [1, 1024, 14]> input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = string("input_735_cast_fp16")];
+            string x_363_pad_type_0 = const()[name = string("x_363_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_363_strides_0 = const()[name = string("x_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_363_pad_0 = const()[name = string("x_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_363_dilations_0 = const()[name = string("x_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_363_groups_0 = const()[name = string("x_363_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277623552))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278672192))))[name = string("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_735_cast_fp16)[name = string("x_363_cast_fp16")];
+            tensor<int32, [3]> input_737_perm_0 = const()[name = string("input_737_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = string("transpose_237")];
+            tensor<fp16, [1, 14, 1024]> input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = string("input_739_cast_fp16")];
+            tensor<int32, [1]> input_741_axes_0 = const()[name = string("input_741_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278674304)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278676416)))];
+            tensor<fp16, [1, 14, 1024]> input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = string("input_741_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278678528))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281824320))))[name = string("encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_13_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281824512)))];
+            tensor<fp16, [1, 14, 4096]> linear_125_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized, x = input_741_cast_fp16)[name = string("linear_125_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = string("input_745_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281832768))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284978560))))[name = string("encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284978752)))];
+            tensor<fp16, [1, 14, 1024]> linear_126_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized, x = input_745_cast_fp16)[name = string("linear_126_cast_fp16")];
+            fp16 var_3401_to_fp16 = const()[name = string("op_3401_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_3402_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3401_to_fp16)[name = string("op_3402_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3402_cast_fp16)[name = string("input_751_cast_fp16")];
+            tensor<int32, [1]> input_753_axes_0 = const()[name = string("input_753_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284980864)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284982976)))];
+            tensor<fp16, [1, 14, 1024]> input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = string("input_753_cast_fp16")];
+            tensor<int32, [4]> cache_57_begin_0 = const()[name = string("cache_57_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_57_end_0 = const()[name = string("cache_57_end_0"), val = tensor<int32, [4]>([15, 1, 42, 1024])];
+            tensor<bool, [4]> cache_57_end_mask_0 = const()[name = string("cache_57_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_57_squeeze_mask_0 = const()[name = string("cache_57_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_57_cast_fp16 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_57_cast_fp16")];
+            tensor<int32, [4]> cache_59_begin_0 = const()[name = string("cache_59_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_59_end_0 = const()[name = string("cache_59_end_0"), val = tensor<int32, [4]>([15, 1, 1024, 8])];
+            tensor<bool, [4]> cache_59_end_mask_0 = const()[name = string("cache_59_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_59_squeeze_mask_0 = const()[name = string("cache_59_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_59_cast_fp16 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_59_cast_fp16")];
+            tensor<int32, [1]> input_755_axes_0 = const()[name = string("input_755_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284985088)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284987200)))];
+            tensor<fp16, [1, 14, 1024]> input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = string("input_755_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284989312))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288135104))))[name = string("encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_14_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288135296)))];
+            tensor<fp16, [1, 14, 4096]> linear_127_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized, x = input_755_cast_fp16)[name = string("linear_127_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = string("input_759_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288143552))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291289344))))[name = string("encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291289536)))];
+            tensor<fp16, [1, 14, 1024]> linear_128_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized, x = input_759_cast_fp16)[name = string("linear_128_cast_fp16")];
+            fp16 var_3438_to_fp16 = const()[name = string("op_3438_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_3439_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3438_to_fp16)[name = string("op_3439_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3439_cast_fp16)[name = string("input_765_cast_fp16")];
+            tensor<int32, [1]> key_29_axes_0 = const()[name = string("key_29_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291291648)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291293760)))];
+            tensor<fp16, [1, 14, 1024]> key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = string("key_29_cast_fp16")];
+            bool input_767_interleave_0 = const()[name = string("input_767_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_767_cast_fp16 = concat(axis = var_69, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = string("input_767_cast_fp16")];
+            tensor<int32, [3]> var_3461_begin_0 = const()[name = string("op_3461_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_3461_end_0 = const()[name = string("op_3461_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3461_end_mask_0 = const()[name = string("op_3461_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_3461_cast_fp16 = slice_by_index(begin = var_3461_begin_0, end = var_3461_end_0, end_mask = var_3461_end_mask_0, x = cache_57_cast_fp16)[name = string("op_3461_cast_fp16")];
+            bool var_3467_interleave_0 = const()[name = string("op_3467_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3467_cast_fp16 = concat(axis = var_69, interleave = var_3467_interleave_0, values = (var_3461_cast_fp16, key_29_cast_fp16))[name = string("op_3467_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291295872))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292082368))))[name = string("encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292082560)))];
+            tensor<fp16, [1, 14, 1024]> linear_129_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized, x = key_29_cast_fp16)[name = string("linear_129_cast_fp16")];
+            tensor<int32, [4]> var_3472 = const()[name = string("op_3472"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_85_cast_fp16 = reshape(shape = var_3472, x = linear_129_cast_fp16)[name = string("q_85_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292084672))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292871168))))[name = string("encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292871360)))];
+            tensor<fp16, [1, 56, 1024]> linear_130_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_130_cast_fp16")];
+            tensor<int32, [4]> var_3477 = const()[name = string("op_3477"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_57_cast_fp16 = reshape(shape = var_3477, x = linear_130_cast_fp16)[name = string("k_57_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292873472))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293659968))))[name = string("encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293660160)))];
+            tensor<fp16, [1, 56, 1024]> linear_131_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_131_cast_fp16")];
+            tensor<int32, [4]> var_3482 = const()[name = string("op_3482"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_29_cast_fp16 = reshape(shape = var_3482, x = linear_131_cast_fp16)[name = string("v_29_cast_fp16")];
+            tensor<int32, [4]> value_37_perm_0 = const()[name = string("value_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293662272)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3495_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = string("op_3495_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293664384)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3497_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = string("op_3497_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_29_perm_0 = const()[name = string("q_with_bias_v_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_371_transpose_x_0 = const()[name = string("x_371_transpose_x_0"), val = bool(false)];
+            bool x_371_transpose_y_0 = const()[name = string("x_371_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_3499_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293666496))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293780224))))[name = string("op_3499_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3497_cast_fp16)[name = string("transpose_236")];
+            tensor<fp16, [1, 8, 14, 111]> x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = op_3499_to_fp16_quantized)[name = string("x_371_cast_fp16")];
+            tensor<int32, [8]> x_373_pad_0 = const()[name = string("x_373_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_373_mode_0 = const()[name = string("x_373_mode_0"), val = string("constant")];
+            fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_373_cast_fp16 = pad(constant_val = const_261_to_fp16, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371_cast_fp16)[name = string("x_373_cast_fp16")];
+            tensor<int32, [4]> var_3507 = const()[name = string("op_3507"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_375_cast_fp16 = reshape(shape = var_3507, x = x_373_cast_fp16)[name = string("x_375_cast_fp16")];
+            tensor<int32, [4]> var_3511_begin_0 = const()[name = string("op_3511_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3511_end_0 = const()[name = string("op_3511_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_3511_end_mask_0 = const()[name = string("op_3511_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_3511_cast_fp16 = slice_by_index(begin = var_3511_begin_0, end = var_3511_end_0, end_mask = var_3511_end_mask_0, x = x_375_cast_fp16)[name = string("op_3511_cast_fp16")];
+            tensor<int32, [4]> var_3512 = const()[name = string("op_3512"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_57_cast_fp16 = reshape(shape = var_3512, x = var_3511_cast_fp16)[name = string("matrix_bd_57_cast_fp16")];
+            bool matrix_ac_29_transpose_x_0 = const()[name = string("matrix_ac_29_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_29_transpose_y_0 = const()[name = string("matrix_ac_29_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_124_perm_0 = const()[name = string("transpose_124_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_125_perm_0 = const()[name = string("transpose_125_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_125 = transpose(perm = transpose_125_perm_0, x = k_57_cast_fp16)[name = string("transpose_234")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_124 = transpose(perm = transpose_124_perm_0, x = var_3495_cast_fp16)[name = string("transpose_235")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_29_cast_fp16 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_124, y = transpose_125)[name = string("matrix_ac_29_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_59_begin_0 = const()[name = string("matrix_bd_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_59_end_0 = const()[name = string("matrix_bd_59_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_59_end_mask_0 = const()[name = string("matrix_bd_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_59_cast_fp16 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57_cast_fp16)[name = string("matrix_bd_59_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3521_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = string("op_3521_cast_fp16")];
+            fp16 _inversed_scores_57_y_0_to_fp16 = const()[name = string("_inversed_scores_57_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_57_cast_fp16 = mul(x = var_3521_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = string("_inversed_scores_57_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_59_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_11)[name = string("scores_59_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3527_cast_fp16 = softmax(axis = var_60, x = scores_59_cast_fp16)[name = string("op_3527_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_769_cast_fp16 = select(a = var_45_to_fp16, b = var_3527_cast_fp16, cond = mask_11)[name = string("input_769_cast_fp16")];
+            bool x_377_transpose_x_0 = const()[name = string("x_377_transpose_x_0"), val = bool(false)];
+            bool x_377_transpose_y_0 = const()[name = string("x_377_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = v_29_cast_fp16)[name = string("transpose_233")];
+            tensor<fp16, [1, 8, 14, 128]> x_377_cast_fp16 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769_cast_fp16, y = value_37_cast_fp16)[name = string("x_377_cast_fp16")];
+            tensor<int32, [4]> var_3531_perm_0 = const()[name = string("op_3531_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3532 = const()[name = string("op_3532"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_3531_cast_fp16 = transpose(perm = var_3531_perm_0, x = x_377_cast_fp16)[name = string("transpose_232")];
+            tensor<fp16, [1, 14, 1024]> input_771_cast_fp16 = reshape(shape = var_3532, x = var_3531_cast_fp16)[name = string("input_771_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293780544))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294567040))))[name = string("encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294567232)))];
+            tensor<fp16, [1, 14, 1024]> linear_133_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized, x = input_771_cast_fp16)[name = string("linear_133_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = string("input_775_cast_fp16")];
+            tensor<int32, [1]> x_381_axes_0 = const()[name = string("x_381_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294569344)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294571456)))];
+            tensor<fp16, [1, 14, 1024]> x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = string("x_381_cast_fp16")];
+            tensor<int32, [3]> input_777_perm_0 = const()[name = string("input_777_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_779_pad_type_0 = const()[name = string("input_779_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_779_strides_0 = const()[name = string("input_779_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_779_pad_0 = const()[name = string("input_779_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_779_dilations_0 = const()[name = string("input_779_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_779_groups_0 = const()[name = string("input_779_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294573568))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296670784))))[name = string("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = string("transpose_231")];
+            tensor<fp16, [1, 2048, 14]> input_779_cast_fp16 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_777_cast_fp16)[name = string("input_779_cast_fp16")];
+            int32 x_383_split_num_splits_0 = const()[name = string("x_383_split_num_splits_0"), val = int32(2)];
+            int32 x_383_split_axis_0 = const()[name = string("x_383_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_383_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_383_split_cast_fp16_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779_cast_fp16)[name = string("x_383_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = string("x_383_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = string("x_383_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_781_cast_fp16 = select(a = var_45_to_fp16, b = x_383_cast_fp16, cond = var_576)[name = string("input_781_cast_fp16")];
+            bool new_x_59_interleave_0 = const()[name = string("new_x_59_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_59_cast_fp16 = concat(axis = var_60, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = string("new_x_59_cast_fp16")];
+            tensor<int32, [3]> var_3571_begin_0 = const()[name = string("op_3571_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_3571_end_0 = const()[name = string("op_3571_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_3571_end_mask_0 = const()[name = string("op_3571_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3571_cast_fp16 = slice_by_index(begin = var_3571_begin_0, end = var_3571_end_0, end_mask = var_3571_end_mask_0, x = new_x_59_cast_fp16)[name = string("op_3571_cast_fp16")];
+            string x_385_pad_type_0 = const()[name = string("x_385_pad_type_0"), val = string("valid")];
+            int32 x_385_groups_0 = const()[name = string("x_385_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_385_strides_0 = const()[name = string("x_385_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_385_pad_0 = const()[name = string("x_385_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_385_dilations_0 = const()[name = string("x_385_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296674944))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296684224))))[name = string("encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_59_cast_fp16)[name = string("x_385_cast_fp16")];
+            tensor<int32, [3]> input_783_perm_0 = const()[name = string("input_783_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_387_axes_0 = const()[name = string("x_387_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296686336)))];
+            tensor<fp16, [1024]> encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296688448)))];
+            tensor<fp16, [1, 14, 1024]> input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = string("transpose_230")];
+            tensor<fp16, [1, 14, 1024]> x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = string("x_387_cast_fp16")];
+            tensor<int32, [3]> input_785_perm_0 = const()[name = string("input_785_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = string("transpose_229")];
+            tensor<fp16, [1, 1024, 14]> input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = string("input_787_cast_fp16")];
+            string x_389_pad_type_0 = const()[name = string("x_389_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_389_strides_0 = const()[name = string("x_389_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_389_pad_0 = const()[name = string("x_389_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_389_dilations_0 = const()[name = string("x_389_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_389_groups_0 = const()[name = string("x_389_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296690560))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297739200))))[name = string("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_389_cast_fp16 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_787_cast_fp16)[name = string("x_389_cast_fp16")];
+            tensor<int32, [3]> input_789_perm_0 = const()[name = string("input_789_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = string("transpose_228")];
+            tensor<fp16, [1, 14, 1024]> input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = string("input_791_cast_fp16")];
+            tensor<int32, [1]> input_793_axes_0 = const()[name = string("input_793_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297741312)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297743424)))];
+            tensor<fp16, [1, 14, 1024]> input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = string("input_793_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297745536))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300891328))))[name = string("encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_14_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300891520)))];
+            tensor<fp16, [1, 14, 4096]> linear_134_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized, x = input_793_cast_fp16)[name = string("linear_134_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = string("input_797_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300899776))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304045568))))[name = string("encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304045760)))];
+            tensor<fp16, [1, 14, 1024]> linear_135_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized, x = input_797_cast_fp16)[name = string("linear_135_cast_fp16")];
+            fp16 var_3614_to_fp16 = const()[name = string("op_3614_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_3615_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3614_to_fp16)[name = string("op_3615_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3615_cast_fp16)[name = string("input_803_cast_fp16")];
+            tensor<int32, [1]> input_805_axes_0 = const()[name = string("input_805_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304047872)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304049984)))];
+            tensor<fp16, [1, 14, 1024]> input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = string("input_805_cast_fp16")];
+            tensor<int32, [4]> cache_61_begin_0 = const()[name = string("cache_61_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_61_end_0 = const()[name = string("cache_61_end_0"), val = tensor<int32, [4]>([16, 1, 42, 1024])];
+            tensor<bool, [4]> cache_61_end_mask_0 = const()[name = string("cache_61_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_61_squeeze_mask_0 = const()[name = string("cache_61_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_61_cast_fp16 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_61_cast_fp16")];
+            tensor<int32, [4]> cache_63_begin_0 = const()[name = string("cache_63_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_63_end_0 = const()[name = string("cache_63_end_0"), val = tensor<int32, [4]>([16, 1, 1024, 8])];
+            tensor<bool, [4]> cache_63_end_mask_0 = const()[name = string("cache_63_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_63_squeeze_mask_0 = const()[name = string("cache_63_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_63_cast_fp16 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_63_cast_fp16")];
+            tensor<int32, [1]> input_807_axes_0 = const()[name = string("input_807_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304052096)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304054208)))];
+            tensor<fp16, [1, 14, 1024]> input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = string("input_807_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304056320))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307202112))))[name = string("encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_15_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307202304)))];
+            tensor<fp16, [1, 14, 4096]> linear_136_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized, x = input_807_cast_fp16)[name = string("linear_136_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = string("input_811_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307210560))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310356352))))[name = string("encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310356544)))];
+            tensor<fp16, [1, 14, 1024]> linear_137_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("linear_137_cast_fp16")];
+            fp16 var_3651_to_fp16 = const()[name = string("op_3651_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_3652_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3651_to_fp16)[name = string("op_3652_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3652_cast_fp16)[name = string("input_817_cast_fp16")];
+            tensor<int32, [1]> key_31_axes_0 = const()[name = string("key_31_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310358656)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310360768)))];
+            tensor<fp16, [1, 14, 1024]> key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = string("key_31_cast_fp16")];
+            bool input_819_interleave_0 = const()[name = string("input_819_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_819_cast_fp16 = concat(axis = var_69, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = string("input_819_cast_fp16")];
+            tensor<int32, [3]> var_3674_begin_0 = const()[name = string("op_3674_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_3674_end_0 = const()[name = string("op_3674_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3674_end_mask_0 = const()[name = string("op_3674_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_3674_cast_fp16 = slice_by_index(begin = var_3674_begin_0, end = var_3674_end_0, end_mask = var_3674_end_mask_0, x = cache_61_cast_fp16)[name = string("op_3674_cast_fp16")];
+            bool var_3680_interleave_0 = const()[name = string("op_3680_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3680_cast_fp16 = concat(axis = var_69, interleave = var_3680_interleave_0, values = (var_3674_cast_fp16, key_31_cast_fp16))[name = string("op_3680_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310362880))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311149376))))[name = string("encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311149568)))];
+            tensor<fp16, [1, 14, 1024]> linear_138_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized, x = key_31_cast_fp16)[name = string("linear_138_cast_fp16")];
+            tensor<int32, [4]> var_3685 = const()[name = string("op_3685"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_91_cast_fp16 = reshape(shape = var_3685, x = linear_138_cast_fp16)[name = string("q_91_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311151680))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311938176))))[name = string("encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311938368)))];
+            tensor<fp16, [1, 56, 1024]> linear_139_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_139_cast_fp16")];
+            tensor<int32, [4]> var_3690 = const()[name = string("op_3690"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_61_cast_fp16 = reshape(shape = var_3690, x = linear_139_cast_fp16)[name = string("k_61_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311940480))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312726976))))[name = string("encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312727168)))];
+            tensor<fp16, [1, 56, 1024]> linear_140_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_140_cast_fp16")];
+            tensor<int32, [4]> var_3695 = const()[name = string("op_3695"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_31_cast_fp16 = reshape(shape = var_3695, x = linear_140_cast_fp16)[name = string("v_31_cast_fp16")];
+            tensor<int32, [4]> value_39_perm_0 = const()[name = string("value_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312729280)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3708_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = string("op_3708_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312731392)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3710_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = string("op_3710_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_31_perm_0 = const()[name = string("q_with_bias_v_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_397_transpose_x_0 = const()[name = string("x_397_transpose_x_0"), val = bool(false)];
+            bool x_397_transpose_y_0 = const()[name = string("x_397_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_3712_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312733504))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312847232))))[name = string("op_3712_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3710_cast_fp16)[name = string("transpose_227")];
+            tensor<fp16, [1, 8, 14, 111]> x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = op_3712_to_fp16_quantized)[name = string("x_397_cast_fp16")];
+            tensor<int32, [8]> x_399_pad_0 = const()[name = string("x_399_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_399_mode_0 = const()[name = string("x_399_mode_0"), val = string("constant")];
+            fp16 const_274_to_fp16 = const()[name = string("const_274_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_399_cast_fp16 = pad(constant_val = const_274_to_fp16, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397_cast_fp16)[name = string("x_399_cast_fp16")];
+            tensor<int32, [4]> var_3720 = const()[name = string("op_3720"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_401_cast_fp16 = reshape(shape = var_3720, x = x_399_cast_fp16)[name = string("x_401_cast_fp16")];
+            tensor<int32, [4]> var_3724_begin_0 = const()[name = string("op_3724_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3724_end_0 = const()[name = string("op_3724_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_3724_end_mask_0 = const()[name = string("op_3724_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_3724_cast_fp16 = slice_by_index(begin = var_3724_begin_0, end = var_3724_end_0, end_mask = var_3724_end_mask_0, x = x_401_cast_fp16)[name = string("op_3724_cast_fp16")];
+            tensor<int32, [4]> var_3725 = const()[name = string("op_3725"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_61_cast_fp16 = reshape(shape = var_3725, x = var_3724_cast_fp16)[name = string("matrix_bd_61_cast_fp16")];
+            bool matrix_ac_31_transpose_x_0 = const()[name = string("matrix_ac_31_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_31_transpose_y_0 = const()[name = string("matrix_ac_31_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_126_perm_0 = const()[name = string("transpose_126_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_127_perm_0 = const()[name = string("transpose_127_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_127 = transpose(perm = transpose_127_perm_0, x = k_61_cast_fp16)[name = string("transpose_225")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_126 = transpose(perm = transpose_126_perm_0, x = var_3708_cast_fp16)[name = string("transpose_226")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_31_cast_fp16 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_126, y = transpose_127)[name = string("matrix_ac_31_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_63_begin_0 = const()[name = string("matrix_bd_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_63_end_0 = const()[name = string("matrix_bd_63_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_63_end_mask_0 = const()[name = string("matrix_bd_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_63_cast_fp16 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61_cast_fp16)[name = string("matrix_bd_63_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3734_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = string("op_3734_cast_fp16")];
+            fp16 _inversed_scores_61_y_0_to_fp16 = const()[name = string("_inversed_scores_61_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_61_cast_fp16 = mul(x = var_3734_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = string("_inversed_scores_61_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_63_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_11)[name = string("scores_63_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3740_cast_fp16 = softmax(axis = var_60, x = scores_63_cast_fp16)[name = string("op_3740_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_821_cast_fp16 = select(a = var_45_to_fp16, b = var_3740_cast_fp16, cond = mask_11)[name = string("input_821_cast_fp16")];
+            bool x_403_transpose_x_0 = const()[name = string("x_403_transpose_x_0"), val = bool(false)];
+            bool x_403_transpose_y_0 = const()[name = string("x_403_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = v_31_cast_fp16)[name = string("transpose_224")];
+            tensor<fp16, [1, 8, 14, 128]> x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821_cast_fp16, y = value_39_cast_fp16)[name = string("x_403_cast_fp16")];
+            tensor<int32, [4]> var_3744_perm_0 = const()[name = string("op_3744_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3745 = const()[name = string("op_3745"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_3744_cast_fp16 = transpose(perm = var_3744_perm_0, x = x_403_cast_fp16)[name = string("transpose_223")];
+            tensor<fp16, [1, 14, 1024]> input_823_cast_fp16 = reshape(shape = var_3745, x = var_3744_cast_fp16)[name = string("input_823_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312847552))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313634048))))[name = string("encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313634240)))];
+            tensor<fp16, [1, 14, 1024]> linear_142_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized, x = input_823_cast_fp16)[name = string("linear_142_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = string("input_827_cast_fp16")];
+            tensor<int32, [1]> x_407_axes_0 = const()[name = string("x_407_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313636352)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313638464)))];
+            tensor<fp16, [1, 14, 1024]> x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = string("x_407_cast_fp16")];
+            tensor<int32, [3]> input_829_perm_0 = const()[name = string("input_829_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_831_pad_type_0 = const()[name = string("input_831_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_831_strides_0 = const()[name = string("input_831_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_831_pad_0 = const()[name = string("input_831_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_831_dilations_0 = const()[name = string("input_831_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_831_groups_0 = const()[name = string("input_831_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313640576))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315737792))))[name = string("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = string("transpose_222")];
+            tensor<fp16, [1, 2048, 14]> input_831_cast_fp16 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_829_cast_fp16)[name = string("input_831_cast_fp16")];
+            int32 x_409_split_num_splits_0 = const()[name = string("x_409_split_num_splits_0"), val = int32(2)];
+            int32 x_409_split_axis_0 = const()[name = string("x_409_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_409_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_409_split_cast_fp16_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831_cast_fp16)[name = string("x_409_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = string("x_409_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = string("x_409_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_833_cast_fp16 = select(a = var_45_to_fp16, b = x_409_cast_fp16, cond = var_576)[name = string("input_833_cast_fp16")];
+            bool new_x_63_interleave_0 = const()[name = string("new_x_63_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_63_cast_fp16 = concat(axis = var_60, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = string("new_x_63_cast_fp16")];
+            tensor<int32, [3]> var_3784_begin_0 = const()[name = string("op_3784_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_3784_end_0 = const()[name = string("op_3784_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_3784_end_mask_0 = const()[name = string("op_3784_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3784_cast_fp16 = slice_by_index(begin = var_3784_begin_0, end = var_3784_end_0, end_mask = var_3784_end_mask_0, x = new_x_63_cast_fp16)[name = string("op_3784_cast_fp16")];
+            string x_411_pad_type_0 = const()[name = string("x_411_pad_type_0"), val = string("valid")];
+            int32 x_411_groups_0 = const()[name = string("x_411_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_411_strides_0 = const()[name = string("x_411_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_411_pad_0 = const()[name = string("x_411_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_411_dilations_0 = const()[name = string("x_411_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315741952))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315751232))))[name = string("encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_411_cast_fp16 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_63_cast_fp16)[name = string("x_411_cast_fp16")];
+            tensor<int32, [3]> input_835_perm_0 = const()[name = string("input_835_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_413_axes_0 = const()[name = string("x_413_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315753344)))];
+            tensor<fp16, [1024]> encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315755456)))];
+            tensor<fp16, [1, 14, 1024]> input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = string("transpose_221")];
+            tensor<fp16, [1, 14, 1024]> x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = string("x_413_cast_fp16")];
+            tensor<int32, [3]> input_837_perm_0 = const()[name = string("input_837_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = string("transpose_220")];
+            tensor<fp16, [1, 1024, 14]> input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = string("input_839_cast_fp16")];
+            string x_415_pad_type_0 = const()[name = string("x_415_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_415_strides_0 = const()[name = string("x_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_415_pad_0 = const()[name = string("x_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_415_dilations_0 = const()[name = string("x_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_415_groups_0 = const()[name = string("x_415_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315757568))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316806208))))[name = string("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_839_cast_fp16)[name = string("x_415_cast_fp16")];
+            tensor<int32, [3]> input_841_perm_0 = const()[name = string("input_841_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = string("transpose_219")];
+            tensor<fp16, [1, 14, 1024]> input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = string("input_843_cast_fp16")];
+            tensor<int32, [1]> input_845_axes_0 = const()[name = string("input_845_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316808320)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316810432)))];
+            tensor<fp16, [1, 14, 1024]> input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = string("input_845_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316812544))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319958336))))[name = string("encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_15_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319958528)))];
+            tensor<fp16, [1, 14, 4096]> linear_143_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized, x = input_845_cast_fp16)[name = string("linear_143_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = string("input_849_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319966784))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323112576))))[name = string("encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323112768)))];
+            tensor<fp16, [1, 14, 1024]> linear_144_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized, x = input_849_cast_fp16)[name = string("linear_144_cast_fp16")];
+            fp16 var_3827_to_fp16 = const()[name = string("op_3827_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_3828_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3827_to_fp16)[name = string("op_3828_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3828_cast_fp16)[name = string("input_855_cast_fp16")];
+            tensor<int32, [1]> input_857_axes_0 = const()[name = string("input_857_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323114880)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323116992)))];
+            tensor<fp16, [1, 14, 1024]> input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = string("input_857_cast_fp16")];
+            tensor<int32, [4]> cache_65_begin_0 = const()[name = string("cache_65_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_65_end_0 = const()[name = string("cache_65_end_0"), val = tensor<int32, [4]>([17, 1, 42, 1024])];
+            tensor<bool, [4]> cache_65_end_mask_0 = const()[name = string("cache_65_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_65_squeeze_mask_0 = const()[name = string("cache_65_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_65_cast_fp16 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_65_cast_fp16")];
+            tensor<int32, [4]> cache_67_begin_0 = const()[name = string("cache_67_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_67_end_0 = const()[name = string("cache_67_end_0"), val = tensor<int32, [4]>([17, 1, 1024, 8])];
+            tensor<bool, [4]> cache_67_end_mask_0 = const()[name = string("cache_67_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_67_squeeze_mask_0 = const()[name = string("cache_67_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_67_cast_fp16 = slice_by_index(begin = cache_67_begin_0, end = cache_67_end_0, end_mask = cache_67_end_mask_0, squeeze_mask = cache_67_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_67_cast_fp16")];
+            tensor<int32, [1]> input_859_axes_0 = const()[name = string("input_859_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323119104)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323121216)))];
+            tensor<fp16, [1, 14, 1024]> input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = string("input_859_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323123328))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326269120))))[name = string("encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_16_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326269312)))];
+            tensor<fp16, [1, 14, 4096]> linear_145_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized, x = input_859_cast_fp16)[name = string("linear_145_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = string("input_863_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326277568))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329423360))))[name = string("encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329423552)))];
+            tensor<fp16, [1, 14, 1024]> linear_146_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized, x = input_863_cast_fp16)[name = string("linear_146_cast_fp16")];
+            fp16 var_3864_to_fp16 = const()[name = string("op_3864_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_3865_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3864_to_fp16)[name = string("op_3865_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3865_cast_fp16)[name = string("input_869_cast_fp16")];
+            tensor<int32, [1]> key_33_axes_0 = const()[name = string("key_33_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329425664)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329427776)))];
+            tensor<fp16, [1, 14, 1024]> key_33_cast_fp16 = layer_norm(axes = key_33_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = string("key_33_cast_fp16")];
+            bool input_871_interleave_0 = const()[name = string("input_871_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_871_cast_fp16 = concat(axis = var_69, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_33_cast_fp16))[name = string("input_871_cast_fp16")];
+            tensor<int32, [3]> var_3887_begin_0 = const()[name = string("op_3887_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_3887_end_0 = const()[name = string("op_3887_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3887_end_mask_0 = const()[name = string("op_3887_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_3887_cast_fp16 = slice_by_index(begin = var_3887_begin_0, end = var_3887_end_0, end_mask = var_3887_end_mask_0, x = cache_65_cast_fp16)[name = string("op_3887_cast_fp16")];
+            bool var_3893_interleave_0 = const()[name = string("op_3893_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3893_cast_fp16 = concat(axis = var_69, interleave = var_3893_interleave_0, values = (var_3887_cast_fp16, key_33_cast_fp16))[name = string("op_3893_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329429888))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330216384))))[name = string("encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330216576)))];
+            tensor<fp16, [1, 14, 1024]> linear_147_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized, x = key_33_cast_fp16)[name = string("linear_147_cast_fp16")];
+            tensor<int32, [4]> var_3898 = const()[name = string("op_3898"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_97_cast_fp16 = reshape(shape = var_3898, x = linear_147_cast_fp16)[name = string("q_97_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330218688))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331005184))))[name = string("encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331005376)))];
+            tensor<fp16, [1, 56, 1024]> linear_148_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_148_cast_fp16")];
+            tensor<int32, [4]> var_3903 = const()[name = string("op_3903"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_65_cast_fp16 = reshape(shape = var_3903, x = linear_148_cast_fp16)[name = string("k_65_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331007488))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331793984))))[name = string("encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331794176)))];
+            tensor<fp16, [1, 56, 1024]> linear_149_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_149_cast_fp16")];
+            tensor<int32, [4]> var_3908 = const()[name = string("op_3908"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_33_cast_fp16 = reshape(shape = var_3908, x = linear_149_cast_fp16)[name = string("v_33_cast_fp16")];
+            tensor<int32, [4]> value_41_perm_0 = const()[name = string("value_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331796288)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3921_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = string("op_3921_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331798400)))];
+            tensor<fp16, [1, 14, 8, 128]> var_3923_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = string("op_3923_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_33_perm_0 = const()[name = string("q_with_bias_v_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_423_transpose_x_0 = const()[name = string("x_423_transpose_x_0"), val = bool(false)];
+            bool x_423_transpose_y_0 = const()[name = string("x_423_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_3925_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331800512))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331914240))))[name = string("op_3925_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_33_cast_fp16 = transpose(perm = q_with_bias_v_33_perm_0, x = var_3923_cast_fp16)[name = string("transpose_218")];
+            tensor<fp16, [1, 8, 14, 111]> x_423_cast_fp16 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_33_cast_fp16, y = op_3925_to_fp16_quantized)[name = string("x_423_cast_fp16")];
+            tensor<int32, [8]> x_425_pad_0 = const()[name = string("x_425_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_425_mode_0 = const()[name = string("x_425_mode_0"), val = string("constant")];
+            fp16 const_287_to_fp16 = const()[name = string("const_287_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_425_cast_fp16 = pad(constant_val = const_287_to_fp16, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423_cast_fp16)[name = string("x_425_cast_fp16")];
+            tensor<int32, [4]> var_3933 = const()[name = string("op_3933"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_427_cast_fp16 = reshape(shape = var_3933, x = x_425_cast_fp16)[name = string("x_427_cast_fp16")];
+            tensor<int32, [4]> var_3937_begin_0 = const()[name = string("op_3937_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3937_end_0 = const()[name = string("op_3937_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_3937_end_mask_0 = const()[name = string("op_3937_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_3937_cast_fp16 = slice_by_index(begin = var_3937_begin_0, end = var_3937_end_0, end_mask = var_3937_end_mask_0, x = x_427_cast_fp16)[name = string("op_3937_cast_fp16")];
+            tensor<int32, [4]> var_3938 = const()[name = string("op_3938"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_65_cast_fp16 = reshape(shape = var_3938, x = var_3937_cast_fp16)[name = string("matrix_bd_65_cast_fp16")];
+            bool matrix_ac_33_transpose_x_0 = const()[name = string("matrix_ac_33_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_33_transpose_y_0 = const()[name = string("matrix_ac_33_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_128_perm_0 = const()[name = string("transpose_128_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_129_perm_0 = const()[name = string("transpose_129_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_129 = transpose(perm = transpose_129_perm_0, x = k_65_cast_fp16)[name = string("transpose_216")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_128 = transpose(perm = transpose_128_perm_0, x = var_3921_cast_fp16)[name = string("transpose_217")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_33_cast_fp16 = matmul(transpose_x = matrix_ac_33_transpose_x_0, transpose_y = matrix_ac_33_transpose_y_0, x = transpose_128, y = transpose_129)[name = string("matrix_ac_33_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_67_begin_0 = const()[name = string("matrix_bd_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_67_end_0 = const()[name = string("matrix_bd_67_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_67_end_mask_0 = const()[name = string("matrix_bd_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_67_cast_fp16 = slice_by_index(begin = matrix_bd_67_begin_0, end = matrix_bd_67_end_0, end_mask = matrix_bd_67_end_mask_0, x = matrix_bd_65_cast_fp16)[name = string("matrix_bd_67_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3947_cast_fp16 = add(x = matrix_ac_33_cast_fp16, y = matrix_bd_67_cast_fp16)[name = string("op_3947_cast_fp16")];
+            fp16 _inversed_scores_65_y_0_to_fp16 = const()[name = string("_inversed_scores_65_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_65_cast_fp16 = mul(x = var_3947_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = string("_inversed_scores_65_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_67_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_11)[name = string("scores_67_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_3953_cast_fp16 = softmax(axis = var_60, x = scores_67_cast_fp16)[name = string("op_3953_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_873_cast_fp16 = select(a = var_45_to_fp16, b = var_3953_cast_fp16, cond = mask_11)[name = string("input_873_cast_fp16")];
+            bool x_429_transpose_x_0 = const()[name = string("x_429_transpose_x_0"), val = bool(false)];
+            bool x_429_transpose_y_0 = const()[name = string("x_429_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = v_33_cast_fp16)[name = string("transpose_215")];
+            tensor<fp16, [1, 8, 14, 128]> x_429_cast_fp16 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873_cast_fp16, y = value_41_cast_fp16)[name = string("x_429_cast_fp16")];
+            tensor<int32, [4]> var_3957_perm_0 = const()[name = string("op_3957_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3958 = const()[name = string("op_3958"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_3957_cast_fp16 = transpose(perm = var_3957_perm_0, x = x_429_cast_fp16)[name = string("transpose_214")];
+            tensor<fp16, [1, 14, 1024]> input_875_cast_fp16 = reshape(shape = var_3958, x = var_3957_cast_fp16)[name = string("input_875_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331914560))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332701056))))[name = string("encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332701248)))];
+            tensor<fp16, [1, 14, 1024]> linear_151_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized, x = input_875_cast_fp16)[name = string("linear_151_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = string("input_879_cast_fp16")];
+            tensor<int32, [1]> x_433_axes_0 = const()[name = string("x_433_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332703360)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332705472)))];
+            tensor<fp16, [1, 14, 1024]> x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = string("x_433_cast_fp16")];
+            tensor<int32, [3]> input_881_perm_0 = const()[name = string("input_881_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_883_pad_type_0 = const()[name = string("input_883_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_883_strides_0 = const()[name = string("input_883_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_883_pad_0 = const()[name = string("input_883_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_883_dilations_0 = const()[name = string("input_883_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_883_groups_0 = const()[name = string("input_883_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332707584))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334804800))))[name = string("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = string("transpose_213")];
+            tensor<fp16, [1, 2048, 14]> input_883_cast_fp16 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_881_cast_fp16)[name = string("input_883_cast_fp16")];
+            int32 x_435_split_num_splits_0 = const()[name = string("x_435_split_num_splits_0"), val = int32(2)];
+            int32 x_435_split_axis_0 = const()[name = string("x_435_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_435_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_435_split_cast_fp16_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883_cast_fp16)[name = string("x_435_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = string("x_435_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = string("x_435_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_885_cast_fp16 = select(a = var_45_to_fp16, b = x_435_cast_fp16, cond = var_576)[name = string("input_885_cast_fp16")];
+            bool new_x_67_interleave_0 = const()[name = string("new_x_67_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_67_cast_fp16 = concat(axis = var_60, interleave = new_x_67_interleave_0, values = (cache_67_cast_fp16, input_885_cast_fp16))[name = string("new_x_67_cast_fp16")];
+            tensor<int32, [3]> var_3997_begin_0 = const()[name = string("op_3997_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_3997_end_0 = const()[name = string("op_3997_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_3997_end_mask_0 = const()[name = string("op_3997_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3997_cast_fp16 = slice_by_index(begin = var_3997_begin_0, end = var_3997_end_0, end_mask = var_3997_end_mask_0, x = new_x_67_cast_fp16)[name = string("op_3997_cast_fp16")];
+            string x_437_pad_type_0 = const()[name = string("x_437_pad_type_0"), val = string("valid")];
+            int32 x_437_groups_0 = const()[name = string("x_437_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_437_strides_0 = const()[name = string("x_437_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_437_pad_0 = const()[name = string("x_437_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_437_dilations_0 = const()[name = string("x_437_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334808960))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334818240))))[name = string("encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_437_cast_fp16 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_67_cast_fp16)[name = string("x_437_cast_fp16")];
+            tensor<int32, [3]> input_887_perm_0 = const()[name = string("input_887_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_439_axes_0 = const()[name = string("x_439_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334820352)))];
+            tensor<fp16, [1024]> encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334822464)))];
+            tensor<fp16, [1, 14, 1024]> input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = string("transpose_212")];
+            tensor<fp16, [1, 14, 1024]> x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = string("x_439_cast_fp16")];
+            tensor<int32, [3]> input_889_perm_0 = const()[name = string("input_889_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = string("transpose_211")];
+            tensor<fp16, [1, 1024, 14]> input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = string("input_891_cast_fp16")];
+            string x_441_pad_type_0 = const()[name = string("x_441_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_441_strides_0 = const()[name = string("x_441_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_441_pad_0 = const()[name = string("x_441_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_441_dilations_0 = const()[name = string("x_441_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_441_groups_0 = const()[name = string("x_441_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334824576))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335873216))))[name = string("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_441_cast_fp16 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_891_cast_fp16)[name = string("x_441_cast_fp16")];
+            tensor<int32, [3]> input_893_perm_0 = const()[name = string("input_893_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = string("transpose_210")];
+            tensor<fp16, [1, 14, 1024]> input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = string("input_895_cast_fp16")];
+            tensor<int32, [1]> input_897_axes_0 = const()[name = string("input_897_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335875328)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335877440)))];
+            tensor<fp16, [1, 14, 1024]> input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = string("input_897_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335879552))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339025344))))[name = string("encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_16_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339025536)))];
+            tensor<fp16, [1, 14, 4096]> linear_152_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized, x = input_897_cast_fp16)[name = string("linear_152_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = string("input_901_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339033792))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342179584))))[name = string("encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342179776)))];
+            tensor<fp16, [1, 14, 1024]> linear_153_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized, x = input_901_cast_fp16)[name = string("linear_153_cast_fp16")];
+            fp16 var_4040_to_fp16 = const()[name = string("op_4040_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4041_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_4040_to_fp16)[name = string("op_4041_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_907_cast_fp16 = add(x = input_895_cast_fp16, y = var_4041_cast_fp16)[name = string("input_907_cast_fp16")];
+            tensor<int32, [1]> input_909_axes_0 = const()[name = string("input_909_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342181888)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342184000)))];
+            tensor<fp16, [1, 14, 1024]> input_909_cast_fp16 = layer_norm(axes = input_909_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_907_cast_fp16)[name = string("input_909_cast_fp16")];
+            tensor<int32, [4]> cache_69_begin_0 = const()[name = string("cache_69_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> cache_69_end_0 = const()[name = string("cache_69_end_0"), val = tensor<int32, [4]>([18, 1, 42, 1024])];
+            tensor<bool, [4]> cache_69_end_mask_0 = const()[name = string("cache_69_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_69_squeeze_mask_0 = const()[name = string("cache_69_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_69_cast_fp16 = slice_by_index(begin = cache_69_begin_0, end = cache_69_end_0, end_mask = cache_69_end_mask_0, squeeze_mask = cache_69_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_69_cast_fp16")];
+            tensor<int32, [4]> cache_71_begin_0 = const()[name = string("cache_71_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> cache_71_end_0 = const()[name = string("cache_71_end_0"), val = tensor<int32, [4]>([18, 1, 1024, 8])];
+            tensor<bool, [4]> cache_71_end_mask_0 = const()[name = string("cache_71_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_71_squeeze_mask_0 = const()[name = string("cache_71_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_71_cast_fp16 = slice_by_index(begin = cache_71_begin_0, end = cache_71_end_0, end_mask = cache_71_end_mask_0, squeeze_mask = cache_71_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_71_cast_fp16")];
+            tensor<int32, [1]> input_911_axes_0 = const()[name = string("input_911_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342186112)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342188224)))];
+            tensor<fp16, [1, 14, 1024]> input_911_cast_fp16 = layer_norm(axes = input_911_axes_0, beta = encoder_layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_feed_forward1_weight_to_fp16, x = input_909_cast_fp16)[name = string("input_911_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342190336))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345336128))))[name = string("encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_17_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345336320)))];
+            tensor<fp16, [1, 14, 4096]> linear_154_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized, x = input_911_cast_fp16)[name = string("linear_154_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_915_cast_fp16 = silu(x = linear_154_cast_fp16)[name = string("input_915_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345344576))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348490368))))[name = string("encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348490560)))];
+            tensor<fp16, [1, 14, 1024]> linear_155_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized, x = input_915_cast_fp16)[name = string("linear_155_cast_fp16")];
+            fp16 var_4077_to_fp16 = const()[name = string("op_4077_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4078_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_4077_to_fp16)[name = string("op_4078_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_921_cast_fp16 = add(x = input_909_cast_fp16, y = var_4078_cast_fp16)[name = string("input_921_cast_fp16")];
+            tensor<int32, [1]> key_35_axes_0 = const()[name = string("key_35_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348492672)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348494784)))];
+            tensor<fp16, [1, 14, 1024]> key_35_cast_fp16 = layer_norm(axes = key_35_axes_0, beta = encoder_layers_17_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_self_att_weight_to_fp16, x = input_921_cast_fp16)[name = string("key_35_cast_fp16")];
+            bool input_923_interleave_0 = const()[name = string("input_923_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_923_cast_fp16 = concat(axis = var_69, interleave = input_923_interleave_0, values = (cache_69_cast_fp16, key_35_cast_fp16))[name = string("input_923_cast_fp16")];
+            tensor<int32, [3]> var_4100_begin_0 = const()[name = string("op_4100_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_4100_end_0 = const()[name = string("op_4100_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4100_end_mask_0 = const()[name = string("op_4100_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_4100_cast_fp16 = slice_by_index(begin = var_4100_begin_0, end = var_4100_end_0, end_mask = var_4100_end_mask_0, x = cache_69_cast_fp16)[name = string("op_4100_cast_fp16")];
+            bool var_4106_interleave_0 = const()[name = string("op_4106_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4106_cast_fp16 = concat(axis = var_69, interleave = var_4106_interleave_0, values = (var_4100_cast_fp16, key_35_cast_fp16))[name = string("op_4106_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348496896))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349283392))))[name = string("encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349283584)))];
+            tensor<fp16, [1, 14, 1024]> linear_156_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized, x = key_35_cast_fp16)[name = string("linear_156_cast_fp16")];
+            tensor<int32, [4]> var_4111 = const()[name = string("op_4111"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_103_cast_fp16 = reshape(shape = var_4111, x = linear_156_cast_fp16)[name = string("q_103_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349285696))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350072192))))[name = string("encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350072384)))];
+            tensor<fp16, [1, 56, 1024]> linear_157_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_157_cast_fp16")];
+            tensor<int32, [4]> var_4116 = const()[name = string("op_4116"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_69_cast_fp16 = reshape(shape = var_4116, x = linear_157_cast_fp16)[name = string("k_69_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350074496))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350860992))))[name = string("encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350861184)))];
+            tensor<fp16, [1, 56, 1024]> linear_158_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_158_cast_fp16")];
+            tensor<int32, [4]> var_4121 = const()[name = string("op_4121"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_35_cast_fp16 = reshape(shape = var_4121, x = linear_158_cast_fp16)[name = string("v_35_cast_fp16")];
+            tensor<int32, [4]> value_43_perm_0 = const()[name = string("value_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_17_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350863296)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4134_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_u_to_fp16)[name = string("op_4134_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_17_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350865408)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4136_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_v_to_fp16)[name = string("op_4136_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_35_perm_0 = const()[name = string("q_with_bias_v_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_449_transpose_x_0 = const()[name = string("x_449_transpose_x_0"), val = bool(false)];
+            bool x_449_transpose_y_0 = const()[name = string("x_449_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_4138_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350867520))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350981248))))[name = string("op_4138_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_35_cast_fp16 = transpose(perm = q_with_bias_v_35_perm_0, x = var_4136_cast_fp16)[name = string("transpose_209")];
+            tensor<fp16, [1, 8, 14, 111]> x_449_cast_fp16 = matmul(transpose_x = x_449_transpose_x_0, transpose_y = x_449_transpose_y_0, x = q_with_bias_v_35_cast_fp16, y = op_4138_to_fp16_quantized)[name = string("x_449_cast_fp16")];
+            tensor<int32, [8]> x_451_pad_0 = const()[name = string("x_451_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_451_mode_0 = const()[name = string("x_451_mode_0"), val = string("constant")];
+            fp16 const_300_to_fp16 = const()[name = string("const_300_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_451_cast_fp16 = pad(constant_val = const_300_to_fp16, mode = x_451_mode_0, pad = x_451_pad_0, x = x_449_cast_fp16)[name = string("x_451_cast_fp16")];
+            tensor<int32, [4]> var_4146 = const()[name = string("op_4146"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_453_cast_fp16 = reshape(shape = var_4146, x = x_451_cast_fp16)[name = string("x_453_cast_fp16")];
+            tensor<int32, [4]> var_4150_begin_0 = const()[name = string("op_4150_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4150_end_0 = const()[name = string("op_4150_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_4150_end_mask_0 = const()[name = string("op_4150_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_4150_cast_fp16 = slice_by_index(begin = var_4150_begin_0, end = var_4150_end_0, end_mask = var_4150_end_mask_0, x = x_453_cast_fp16)[name = string("op_4150_cast_fp16")];
+            tensor<int32, [4]> var_4151 = const()[name = string("op_4151"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_69_cast_fp16 = reshape(shape = var_4151, x = var_4150_cast_fp16)[name = string("matrix_bd_69_cast_fp16")];
+            bool matrix_ac_35_transpose_x_0 = const()[name = string("matrix_ac_35_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_35_transpose_y_0 = const()[name = string("matrix_ac_35_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_130_perm_0 = const()[name = string("transpose_130_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_131_perm_0 = const()[name = string("transpose_131_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_131 = transpose(perm = transpose_131_perm_0, x = k_69_cast_fp16)[name = string("transpose_207")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_130 = transpose(perm = transpose_130_perm_0, x = var_4134_cast_fp16)[name = string("transpose_208")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_35_cast_fp16 = matmul(transpose_x = matrix_ac_35_transpose_x_0, transpose_y = matrix_ac_35_transpose_y_0, x = transpose_130, y = transpose_131)[name = string("matrix_ac_35_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_71_begin_0 = const()[name = string("matrix_bd_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_71_end_0 = const()[name = string("matrix_bd_71_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_71_end_mask_0 = const()[name = string("matrix_bd_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_71_cast_fp16 = slice_by_index(begin = matrix_bd_71_begin_0, end = matrix_bd_71_end_0, end_mask = matrix_bd_71_end_mask_0, x = matrix_bd_69_cast_fp16)[name = string("matrix_bd_71_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_4160_cast_fp16 = add(x = matrix_ac_35_cast_fp16, y = matrix_bd_71_cast_fp16)[name = string("op_4160_cast_fp16")];
+            fp16 _inversed_scores_69_y_0_to_fp16 = const()[name = string("_inversed_scores_69_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_69_cast_fp16 = mul(x = var_4160_cast_fp16, y = _inversed_scores_69_y_0_to_fp16)[name = string("_inversed_scores_69_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_71_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_69_cast_fp16, cond = mask_11)[name = string("scores_71_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_4166_cast_fp16 = softmax(axis = var_60, x = scores_71_cast_fp16)[name = string("op_4166_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_925_cast_fp16 = select(a = var_45_to_fp16, b = var_4166_cast_fp16, cond = mask_11)[name = string("input_925_cast_fp16")];
+            bool x_455_transpose_x_0 = const()[name = string("x_455_transpose_x_0"), val = bool(false)];
+            bool x_455_transpose_y_0 = const()[name = string("x_455_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = v_35_cast_fp16)[name = string("transpose_206")];
+            tensor<fp16, [1, 8, 14, 128]> x_455_cast_fp16 = matmul(transpose_x = x_455_transpose_x_0, transpose_y = x_455_transpose_y_0, x = input_925_cast_fp16, y = value_43_cast_fp16)[name = string("x_455_cast_fp16")];
+            tensor<int32, [4]> var_4170_perm_0 = const()[name = string("op_4170_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4171 = const()[name = string("op_4171"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_4170_cast_fp16 = transpose(perm = var_4170_perm_0, x = x_455_cast_fp16)[name = string("transpose_205")];
+            tensor<fp16, [1, 14, 1024]> input_927_cast_fp16 = reshape(shape = var_4171, x = var_4170_cast_fp16)[name = string("input_927_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350981568))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351768064))))[name = string("encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351768256)))];
+            tensor<fp16, [1, 14, 1024]> linear_160_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized, x = input_927_cast_fp16)[name = string("linear_160_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_931_cast_fp16 = add(x = input_921_cast_fp16, y = linear_160_cast_fp16)[name = string("input_931_cast_fp16")];
+            tensor<int32, [1]> x_459_axes_0 = const()[name = string("x_459_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351770368)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351772480)))];
+            tensor<fp16, [1, 14, 1024]> x_459_cast_fp16 = layer_norm(axes = x_459_axes_0, beta = encoder_layers_17_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_conv_weight_to_fp16, x = input_931_cast_fp16)[name = string("x_459_cast_fp16")];
+            tensor<int32, [3]> input_933_perm_0 = const()[name = string("input_933_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_935_pad_type_0 = const()[name = string("input_935_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_935_strides_0 = const()[name = string("input_935_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_935_pad_0 = const()[name = string("input_935_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_935_dilations_0 = const()[name = string("input_935_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_935_groups_0 = const()[name = string("input_935_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351774592))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353871808))))[name = string("encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_933_cast_fp16 = transpose(perm = input_933_perm_0, x = x_459_cast_fp16)[name = string("transpose_204")];
+            tensor<fp16, [1, 2048, 14]> input_935_cast_fp16 = conv(dilations = input_935_dilations_0, groups = input_935_groups_0, pad = input_935_pad_0, pad_type = input_935_pad_type_0, strides = input_935_strides_0, weight = encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_933_cast_fp16)[name = string("input_935_cast_fp16")];
+            int32 x_461_split_num_splits_0 = const()[name = string("x_461_split_num_splits_0"), val = int32(2)];
+            int32 x_461_split_axis_0 = const()[name = string("x_461_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_461_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_461_split_cast_fp16_1 = split(axis = x_461_split_axis_0, num_splits = x_461_split_num_splits_0, x = input_935_cast_fp16)[name = string("x_461_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_461_split_1_sigmoid_cast_fp16 = sigmoid(x = x_461_split_cast_fp16_1)[name = string("x_461_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_461_cast_fp16 = mul(x = x_461_split_cast_fp16_0, y = x_461_split_1_sigmoid_cast_fp16)[name = string("x_461_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_937_cast_fp16 = select(a = var_45_to_fp16, b = x_461_cast_fp16, cond = var_576)[name = string("input_937_cast_fp16")];
+            bool new_x_71_interleave_0 = const()[name = string("new_x_71_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_71_cast_fp16 = concat(axis = var_60, interleave = new_x_71_interleave_0, values = (cache_71_cast_fp16, input_937_cast_fp16))[name = string("new_x_71_cast_fp16")];
+            tensor<int32, [3]> var_4210_begin_0 = const()[name = string("op_4210_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_4210_end_0 = const()[name = string("op_4210_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_4210_end_mask_0 = const()[name = string("op_4210_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4210_cast_fp16 = slice_by_index(begin = var_4210_begin_0, end = var_4210_end_0, end_mask = var_4210_end_mask_0, x = new_x_71_cast_fp16)[name = string("op_4210_cast_fp16")];
+            string x_463_pad_type_0 = const()[name = string("x_463_pad_type_0"), val = string("valid")];
+            int32 x_463_groups_0 = const()[name = string("x_463_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_463_strides_0 = const()[name = string("x_463_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_463_pad_0 = const()[name = string("x_463_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_463_dilations_0 = const()[name = string("x_463_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353875968))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353885248))))[name = string("encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_463_cast_fp16 = conv(dilations = x_463_dilations_0, groups = x_463_groups_0, pad = x_463_pad_0, pad_type = x_463_pad_type_0, strides = x_463_strides_0, weight = encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_71_cast_fp16)[name = string("x_463_cast_fp16")];
+            tensor<int32, [3]> input_939_perm_0 = const()[name = string("input_939_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_465_axes_0 = const()[name = string("x_465_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353887360)))];
+            tensor<fp16, [1024]> encoder_layers_17_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353889472)))];
+            tensor<fp16, [1, 14, 1024]> input_939_cast_fp16 = transpose(perm = input_939_perm_0, x = x_463_cast_fp16)[name = string("transpose_203")];
+            tensor<fp16, [1, 14, 1024]> x_465_cast_fp16 = layer_norm(axes = x_465_axes_0, beta = encoder_layers_17_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_conv_batch_norm_weight_to_fp16, x = input_939_cast_fp16)[name = string("x_465_cast_fp16")];
+            tensor<int32, [3]> input_941_perm_0 = const()[name = string("input_941_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_941_cast_fp16 = transpose(perm = input_941_perm_0, x = x_465_cast_fp16)[name = string("transpose_202")];
+            tensor<fp16, [1, 1024, 14]> input_943_cast_fp16 = silu(x = input_941_cast_fp16)[name = string("input_943_cast_fp16")];
+            string x_467_pad_type_0 = const()[name = string("x_467_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_467_strides_0 = const()[name = string("x_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_467_pad_0 = const()[name = string("x_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_467_dilations_0 = const()[name = string("x_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_467_groups_0 = const()[name = string("x_467_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353891584))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354940224))))[name = string("encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_467_cast_fp16 = conv(dilations = x_467_dilations_0, groups = x_467_groups_0, pad = x_467_pad_0, pad_type = x_467_pad_type_0, strides = x_467_strides_0, weight = encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_943_cast_fp16)[name = string("x_467_cast_fp16")];
+            tensor<int32, [3]> input_945_perm_0 = const()[name = string("input_945_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_945_cast_fp16 = transpose(perm = input_945_perm_0, x = x_467_cast_fp16)[name = string("transpose_201")];
+            tensor<fp16, [1, 14, 1024]> input_947_cast_fp16 = add(x = input_931_cast_fp16, y = input_945_cast_fp16)[name = string("input_947_cast_fp16")];
+            tensor<int32, [1]> input_949_axes_0 = const()[name = string("input_949_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354942336)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354944448)))];
+            tensor<fp16, [1, 14, 1024]> input_949_cast_fp16 = layer_norm(axes = input_949_axes_0, beta = encoder_layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_feed_forward2_weight_to_fp16, x = input_947_cast_fp16)[name = string("input_949_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354946560))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358092352))))[name = string("encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_17_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358092544)))];
+            tensor<fp16, [1, 14, 4096]> linear_161_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("linear_161_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_953_cast_fp16 = silu(x = linear_161_cast_fp16)[name = string("input_953_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358100800))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361246592))))[name = string("encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361246784)))];
+            tensor<fp16, [1, 14, 1024]> linear_162_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized, x = input_953_cast_fp16)[name = string("linear_162_cast_fp16")];
+            fp16 var_4253_to_fp16 = const()[name = string("op_4253_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4254_cast_fp16 = mul(x = linear_162_cast_fp16, y = var_4253_to_fp16)[name = string("op_4254_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_959_cast_fp16 = add(x = input_947_cast_fp16, y = var_4254_cast_fp16)[name = string("input_959_cast_fp16")];
+            tensor<int32, [1]> input_961_axes_0 = const()[name = string("input_961_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361248896)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361251008)))];
+            tensor<fp16, [1, 14, 1024]> input_961_cast_fp16 = layer_norm(axes = input_961_axes_0, beta = encoder_layers_17_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_17_norm_out_weight_to_fp16, x = input_959_cast_fp16)[name = string("input_961_cast_fp16")];
+            tensor<int32, [4]> cache_73_begin_0 = const()[name = string("cache_73_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> cache_73_end_0 = const()[name = string("cache_73_end_0"), val = tensor<int32, [4]>([19, 1, 42, 1024])];
+            tensor<bool, [4]> cache_73_end_mask_0 = const()[name = string("cache_73_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_73_squeeze_mask_0 = const()[name = string("cache_73_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_73_cast_fp16 = slice_by_index(begin = cache_73_begin_0, end = cache_73_end_0, end_mask = cache_73_end_mask_0, squeeze_mask = cache_73_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_73_cast_fp16")];
+            tensor<int32, [4]> cache_75_begin_0 = const()[name = string("cache_75_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> cache_75_end_0 = const()[name = string("cache_75_end_0"), val = tensor<int32, [4]>([19, 1, 1024, 8])];
+            tensor<bool, [4]> cache_75_end_mask_0 = const()[name = string("cache_75_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_75_squeeze_mask_0 = const()[name = string("cache_75_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_75_cast_fp16 = slice_by_index(begin = cache_75_begin_0, end = cache_75_end_0, end_mask = cache_75_end_mask_0, squeeze_mask = cache_75_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_75_cast_fp16")];
+            tensor<int32, [1]> input_963_axes_0 = const()[name = string("input_963_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361253120)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361255232)))];
+            tensor<fp16, [1, 14, 1024]> input_963_cast_fp16 = layer_norm(axes = input_963_axes_0, beta = encoder_layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_feed_forward1_weight_to_fp16, x = input_961_cast_fp16)[name = string("input_963_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361257344))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364403136))))[name = string("encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_18_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364403328)))];
+            tensor<fp16, [1, 14, 4096]> linear_163_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized, x = input_963_cast_fp16)[name = string("linear_163_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_967_cast_fp16 = silu(x = linear_163_cast_fp16)[name = string("input_967_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364411584))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367557376))))[name = string("encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367557568)))];
+            tensor<fp16, [1, 14, 1024]> linear_164_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized, x = input_967_cast_fp16)[name = string("linear_164_cast_fp16")];
+            fp16 var_4290_to_fp16 = const()[name = string("op_4290_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4291_cast_fp16 = mul(x = linear_164_cast_fp16, y = var_4290_to_fp16)[name = string("op_4291_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_973_cast_fp16 = add(x = input_961_cast_fp16, y = var_4291_cast_fp16)[name = string("input_973_cast_fp16")];
+            tensor<int32, [1]> key_37_axes_0 = const()[name = string("key_37_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367559680)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367561792)))];
+            tensor<fp16, [1, 14, 1024]> key_37_cast_fp16 = layer_norm(axes = key_37_axes_0, beta = encoder_layers_18_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_self_att_weight_to_fp16, x = input_973_cast_fp16)[name = string("key_37_cast_fp16")];
+            bool input_975_interleave_0 = const()[name = string("input_975_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_975_cast_fp16 = concat(axis = var_69, interleave = input_975_interleave_0, values = (cache_73_cast_fp16, key_37_cast_fp16))[name = string("input_975_cast_fp16")];
+            tensor<int32, [3]> var_4313_begin_0 = const()[name = string("op_4313_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_4313_end_0 = const()[name = string("op_4313_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4313_end_mask_0 = const()[name = string("op_4313_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_4313_cast_fp16 = slice_by_index(begin = var_4313_begin_0, end = var_4313_end_0, end_mask = var_4313_end_mask_0, x = cache_73_cast_fp16)[name = string("op_4313_cast_fp16")];
+            bool var_4319_interleave_0 = const()[name = string("op_4319_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4319_cast_fp16 = concat(axis = var_69, interleave = var_4319_interleave_0, values = (var_4313_cast_fp16, key_37_cast_fp16))[name = string("op_4319_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367563904))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368350400))))[name = string("encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368350592)))];
+            tensor<fp16, [1, 14, 1024]> linear_165_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized, x = key_37_cast_fp16)[name = string("linear_165_cast_fp16")];
+            tensor<int32, [4]> var_4324 = const()[name = string("op_4324"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_109_cast_fp16 = reshape(shape = var_4324, x = linear_165_cast_fp16)[name = string("q_109_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368352704))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369139200))))[name = string("encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369139392)))];
+            tensor<fp16, [1, 56, 1024]> linear_166_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_166_cast_fp16")];
+            tensor<int32, [4]> var_4329 = const()[name = string("op_4329"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_73_cast_fp16 = reshape(shape = var_4329, x = linear_166_cast_fp16)[name = string("k_73_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369141504))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369928000))))[name = string("encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369928192)))];
+            tensor<fp16, [1, 56, 1024]> linear_167_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_167_cast_fp16")];
+            tensor<int32, [4]> var_4334 = const()[name = string("op_4334"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_37_cast_fp16 = reshape(shape = var_4334, x = linear_167_cast_fp16)[name = string("v_37_cast_fp16")];
+            tensor<int32, [4]> value_45_perm_0 = const()[name = string("value_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_18_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369930304)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4347_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_u_to_fp16)[name = string("op_4347_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_18_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369932416)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4349_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_v_to_fp16)[name = string("op_4349_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_37_perm_0 = const()[name = string("q_with_bias_v_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_475_transpose_x_0 = const()[name = string("x_475_transpose_x_0"), val = bool(false)];
+            bool x_475_transpose_y_0 = const()[name = string("x_475_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_4351_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369934528))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370048256))))[name = string("op_4351_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_37_cast_fp16 = transpose(perm = q_with_bias_v_37_perm_0, x = var_4349_cast_fp16)[name = string("transpose_200")];
+            tensor<fp16, [1, 8, 14, 111]> x_475_cast_fp16 = matmul(transpose_x = x_475_transpose_x_0, transpose_y = x_475_transpose_y_0, x = q_with_bias_v_37_cast_fp16, y = op_4351_to_fp16_quantized)[name = string("x_475_cast_fp16")];
+            tensor<int32, [8]> x_477_pad_0 = const()[name = string("x_477_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_477_mode_0 = const()[name = string("x_477_mode_0"), val = string("constant")];
+            fp16 const_313_to_fp16 = const()[name = string("const_313_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_477_cast_fp16 = pad(constant_val = const_313_to_fp16, mode = x_477_mode_0, pad = x_477_pad_0, x = x_475_cast_fp16)[name = string("x_477_cast_fp16")];
+            tensor<int32, [4]> var_4359 = const()[name = string("op_4359"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_479_cast_fp16 = reshape(shape = var_4359, x = x_477_cast_fp16)[name = string("x_479_cast_fp16")];
+            tensor<int32, [4]> var_4363_begin_0 = const()[name = string("op_4363_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4363_end_0 = const()[name = string("op_4363_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_4363_end_mask_0 = const()[name = string("op_4363_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_4363_cast_fp16 = slice_by_index(begin = var_4363_begin_0, end = var_4363_end_0, end_mask = var_4363_end_mask_0, x = x_479_cast_fp16)[name = string("op_4363_cast_fp16")];
+            tensor<int32, [4]> var_4364 = const()[name = string("op_4364"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_73_cast_fp16 = reshape(shape = var_4364, x = var_4363_cast_fp16)[name = string("matrix_bd_73_cast_fp16")];
+            bool matrix_ac_37_transpose_x_0 = const()[name = string("matrix_ac_37_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_37_transpose_y_0 = const()[name = string("matrix_ac_37_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_132_perm_0 = const()[name = string("transpose_132_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_133_perm_0 = const()[name = string("transpose_133_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_133 = transpose(perm = transpose_133_perm_0, x = k_73_cast_fp16)[name = string("transpose_198")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_132 = transpose(perm = transpose_132_perm_0, x = var_4347_cast_fp16)[name = string("transpose_199")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_37_cast_fp16 = matmul(transpose_x = matrix_ac_37_transpose_x_0, transpose_y = matrix_ac_37_transpose_y_0, x = transpose_132, y = transpose_133)[name = string("matrix_ac_37_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_75_begin_0 = const()[name = string("matrix_bd_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_75_end_0 = const()[name = string("matrix_bd_75_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_75_end_mask_0 = const()[name = string("matrix_bd_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_75_cast_fp16 = slice_by_index(begin = matrix_bd_75_begin_0, end = matrix_bd_75_end_0, end_mask = matrix_bd_75_end_mask_0, x = matrix_bd_73_cast_fp16)[name = string("matrix_bd_75_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_4373_cast_fp16 = add(x = matrix_ac_37_cast_fp16, y = matrix_bd_75_cast_fp16)[name = string("op_4373_cast_fp16")];
+            fp16 _inversed_scores_73_y_0_to_fp16 = const()[name = string("_inversed_scores_73_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_73_cast_fp16 = mul(x = var_4373_cast_fp16, y = _inversed_scores_73_y_0_to_fp16)[name = string("_inversed_scores_73_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_75_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_73_cast_fp16, cond = mask_11)[name = string("scores_75_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_4379_cast_fp16 = softmax(axis = var_60, x = scores_75_cast_fp16)[name = string("op_4379_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_977_cast_fp16 = select(a = var_45_to_fp16, b = var_4379_cast_fp16, cond = mask_11)[name = string("input_977_cast_fp16")];
+            bool x_481_transpose_x_0 = const()[name = string("x_481_transpose_x_0"), val = bool(false)];
+            bool x_481_transpose_y_0 = const()[name = string("x_481_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = v_37_cast_fp16)[name = string("transpose_197")];
+            tensor<fp16, [1, 8, 14, 128]> x_481_cast_fp16 = matmul(transpose_x = x_481_transpose_x_0, transpose_y = x_481_transpose_y_0, x = input_977_cast_fp16, y = value_45_cast_fp16)[name = string("x_481_cast_fp16")];
+            tensor<int32, [4]> var_4383_perm_0 = const()[name = string("op_4383_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4384 = const()[name = string("op_4384"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_4383_cast_fp16 = transpose(perm = var_4383_perm_0, x = x_481_cast_fp16)[name = string("transpose_196")];
+            tensor<fp16, [1, 14, 1024]> input_979_cast_fp16 = reshape(shape = var_4384, x = var_4383_cast_fp16)[name = string("input_979_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370048576))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371097216))))[name = string("encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371099328)))];
+            tensor<fp16, [1, 14, 1024]> linear_169_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized, x = input_979_cast_fp16)[name = string("linear_169_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_983_cast_fp16 = add(x = input_973_cast_fp16, y = linear_169_cast_fp16)[name = string("input_983_cast_fp16")];
+            tensor<int32, [1]> x_485_axes_0 = const()[name = string("x_485_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371101440)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371103552)))];
+            tensor<fp16, [1, 14, 1024]> x_485_cast_fp16 = layer_norm(axes = x_485_axes_0, beta = encoder_layers_18_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_conv_weight_to_fp16, x = input_983_cast_fp16)[name = string("x_485_cast_fp16")];
+            tensor<int32, [3]> input_985_perm_0 = const()[name = string("input_985_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_987_pad_type_0 = const()[name = string("input_987_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_987_strides_0 = const()[name = string("input_987_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_987_pad_0 = const()[name = string("input_987_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_987_dilations_0 = const()[name = string("input_987_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_987_groups_0 = const()[name = string("input_987_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371105664))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373202880))))[name = string("encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_985_cast_fp16 = transpose(perm = input_985_perm_0, x = x_485_cast_fp16)[name = string("transpose_195")];
+            tensor<fp16, [1, 2048, 14]> input_987_cast_fp16 = conv(dilations = input_987_dilations_0, groups = input_987_groups_0, pad = input_987_pad_0, pad_type = input_987_pad_type_0, strides = input_987_strides_0, weight = encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_985_cast_fp16)[name = string("input_987_cast_fp16")];
+            int32 x_487_split_num_splits_0 = const()[name = string("x_487_split_num_splits_0"), val = int32(2)];
+            int32 x_487_split_axis_0 = const()[name = string("x_487_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_487_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_487_split_cast_fp16_1 = split(axis = x_487_split_axis_0, num_splits = x_487_split_num_splits_0, x = input_987_cast_fp16)[name = string("x_487_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_487_split_1_sigmoid_cast_fp16 = sigmoid(x = x_487_split_cast_fp16_1)[name = string("x_487_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_487_cast_fp16 = mul(x = x_487_split_cast_fp16_0, y = x_487_split_1_sigmoid_cast_fp16)[name = string("x_487_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_989_cast_fp16 = select(a = var_45_to_fp16, b = x_487_cast_fp16, cond = var_576)[name = string("input_989_cast_fp16")];
+            bool new_x_75_interleave_0 = const()[name = string("new_x_75_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_75_cast_fp16 = concat(axis = var_60, interleave = new_x_75_interleave_0, values = (cache_75_cast_fp16, input_989_cast_fp16))[name = string("new_x_75_cast_fp16")];
+            tensor<int32, [3]> var_4423_begin_0 = const()[name = string("op_4423_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_4423_end_0 = const()[name = string("op_4423_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_4423_end_mask_0 = const()[name = string("op_4423_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4423_cast_fp16 = slice_by_index(begin = var_4423_begin_0, end = var_4423_end_0, end_mask = var_4423_end_mask_0, x = new_x_75_cast_fp16)[name = string("op_4423_cast_fp16")];
+            string x_489_pad_type_0 = const()[name = string("x_489_pad_type_0"), val = string("valid")];
+            int32 x_489_groups_0 = const()[name = string("x_489_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_489_strides_0 = const()[name = string("x_489_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_489_pad_0 = const()[name = string("x_489_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_489_dilations_0 = const()[name = string("x_489_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373207040))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373216320))))[name = string("encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_489_cast_fp16 = conv(dilations = x_489_dilations_0, groups = x_489_groups_0, pad = x_489_pad_0, pad_type = x_489_pad_type_0, strides = x_489_strides_0, weight = encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_75_cast_fp16)[name = string("x_489_cast_fp16")];
+            tensor<int32, [3]> input_991_perm_0 = const()[name = string("input_991_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_491_axes_0 = const()[name = string("x_491_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373218432)))];
+            tensor<fp16, [1024]> encoder_layers_18_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373220544)))];
+            tensor<fp16, [1, 14, 1024]> input_991_cast_fp16 = transpose(perm = input_991_perm_0, x = x_489_cast_fp16)[name = string("transpose_194")];
+            tensor<fp16, [1, 14, 1024]> x_491_cast_fp16 = layer_norm(axes = x_491_axes_0, beta = encoder_layers_18_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_conv_batch_norm_weight_to_fp16, x = input_991_cast_fp16)[name = string("x_491_cast_fp16")];
+            tensor<int32, [3]> input_993_perm_0 = const()[name = string("input_993_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_993_cast_fp16 = transpose(perm = input_993_perm_0, x = x_491_cast_fp16)[name = string("transpose_193")];
+            tensor<fp16, [1, 1024, 14]> input_995_cast_fp16 = silu(x = input_993_cast_fp16)[name = string("input_995_cast_fp16")];
+            string x_493_pad_type_0 = const()[name = string("x_493_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_493_strides_0 = const()[name = string("x_493_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_493_pad_0 = const()[name = string("x_493_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_493_dilations_0 = const()[name = string("x_493_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_493_groups_0 = const()[name = string("x_493_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373222656))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374271296))))[name = string("encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_493_cast_fp16 = conv(dilations = x_493_dilations_0, groups = x_493_groups_0, pad = x_493_pad_0, pad_type = x_493_pad_type_0, strides = x_493_strides_0, weight = encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_995_cast_fp16)[name = string("x_493_cast_fp16")];
+            tensor<int32, [3]> input_997_perm_0 = const()[name = string("input_997_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_997_cast_fp16 = transpose(perm = input_997_perm_0, x = x_493_cast_fp16)[name = string("transpose_192")];
+            tensor<fp16, [1, 14, 1024]> input_999_cast_fp16 = add(x = input_983_cast_fp16, y = input_997_cast_fp16)[name = string("input_999_cast_fp16")];
+            tensor<int32, [1]> input_1001_axes_0 = const()[name = string("input_1001_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374273408)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374275520)))];
+            tensor<fp16, [1, 14, 1024]> input_1001_cast_fp16 = layer_norm(axes = input_1001_axes_0, beta = encoder_layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_feed_forward2_weight_to_fp16, x = input_999_cast_fp16)[name = string("input_1001_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374277632))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378472000))))[name = string("encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_18_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378480256)))];
+            tensor<fp16, [1, 14, 4096]> linear_170_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1001_cast_fp16)[name = string("linear_170_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1005_cast_fp16 = silu(x = linear_170_cast_fp16)[name = string("input_1005_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378488512))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382682880))))[name = string("encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_18_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382684992)))];
+            tensor<fp16, [1, 14, 1024]> linear_171_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1005_cast_fp16)[name = string("linear_171_cast_fp16")];
+            fp16 var_4466_to_fp16 = const()[name = string("op_4466_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4467_cast_fp16 = mul(x = linear_171_cast_fp16, y = var_4466_to_fp16)[name = string("op_4467_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1011_cast_fp16 = add(x = input_999_cast_fp16, y = var_4467_cast_fp16)[name = string("input_1011_cast_fp16")];
+            tensor<int32, [1]> input_1013_axes_0 = const()[name = string("input_1013_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382687104)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382689216)))];
+            tensor<fp16, [1, 14, 1024]> input_1013_cast_fp16 = layer_norm(axes = input_1013_axes_0, beta = encoder_layers_18_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_18_norm_out_weight_to_fp16, x = input_1011_cast_fp16)[name = string("input_1013_cast_fp16")];
+            tensor<int32, [4]> cache_77_begin_0 = const()[name = string("cache_77_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
+            tensor<int32, [4]> cache_77_end_0 = const()[name = string("cache_77_end_0"), val = tensor<int32, [4]>([20, 1, 42, 1024])];
+            tensor<bool, [4]> cache_77_end_mask_0 = const()[name = string("cache_77_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_77_squeeze_mask_0 = const()[name = string("cache_77_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_77_cast_fp16 = slice_by_index(begin = cache_77_begin_0, end = cache_77_end_0, end_mask = cache_77_end_mask_0, squeeze_mask = cache_77_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_77_cast_fp16")];
+            tensor<int32, [4]> cache_79_begin_0 = const()[name = string("cache_79_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
+            tensor<int32, [4]> cache_79_end_0 = const()[name = string("cache_79_end_0"), val = tensor<int32, [4]>([20, 1, 1024, 8])];
+            tensor<bool, [4]> cache_79_end_mask_0 = const()[name = string("cache_79_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_79_squeeze_mask_0 = const()[name = string("cache_79_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_79_cast_fp16 = slice_by_index(begin = cache_79_begin_0, end = cache_79_end_0, end_mask = cache_79_end_mask_0, squeeze_mask = cache_79_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_79_cast_fp16")];
+            tensor<int32, [1]> input_1015_axes_0 = const()[name = string("input_1015_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382691328)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382693440)))];
+            tensor<fp16, [1, 14, 1024]> input_1015_cast_fp16 = layer_norm(axes = input_1015_axes_0, beta = encoder_layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_feed_forward1_weight_to_fp16, x = input_1013_cast_fp16)[name = string("input_1015_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382695552))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386889920))))[name = string("encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_19_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386898176)))];
+            tensor<fp16, [1, 14, 4096]> linear_172_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1015_cast_fp16)[name = string("linear_172_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1019_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("input_1019_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386906432))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391100800))))[name = string("encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391102912)))];
+            tensor<fp16, [1, 14, 1024]> linear_173_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1019_cast_fp16)[name = string("linear_173_cast_fp16")];
+            fp16 var_4503_to_fp16 = const()[name = string("op_4503_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4504_cast_fp16 = mul(x = linear_173_cast_fp16, y = var_4503_to_fp16)[name = string("op_4504_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1025_cast_fp16 = add(x = input_1013_cast_fp16, y = var_4504_cast_fp16)[name = string("input_1025_cast_fp16")];
+            tensor<int32, [1]> key_39_axes_0 = const()[name = string("key_39_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391105024)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391107136)))];
+            tensor<fp16, [1, 14, 1024]> key_39_cast_fp16 = layer_norm(axes = key_39_axes_0, beta = encoder_layers_19_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_self_att_weight_to_fp16, x = input_1025_cast_fp16)[name = string("key_39_cast_fp16")];
+            bool input_1027_interleave_0 = const()[name = string("input_1027_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_1027_cast_fp16 = concat(axis = var_69, interleave = input_1027_interleave_0, values = (cache_77_cast_fp16, key_39_cast_fp16))[name = string("input_1027_cast_fp16")];
+            tensor<int32, [3]> var_4526_begin_0 = const()[name = string("op_4526_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_4526_end_0 = const()[name = string("op_4526_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4526_end_mask_0 = const()[name = string("op_4526_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_4526_cast_fp16 = slice_by_index(begin = var_4526_begin_0, end = var_4526_end_0, end_mask = var_4526_end_mask_0, x = cache_77_cast_fp16)[name = string("op_4526_cast_fp16")];
+            bool var_4532_interleave_0 = const()[name = string("op_4532_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4532_cast_fp16 = concat(axis = var_69, interleave = var_4532_interleave_0, values = (var_4526_cast_fp16, key_39_cast_fp16))[name = string("op_4532_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391109248))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392157888))))[name = string("encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392160000)))];
+            tensor<fp16, [1, 14, 1024]> linear_174_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized, x = key_39_cast_fp16)[name = string("linear_174_cast_fp16")];
+            tensor<int32, [4]> var_4537 = const()[name = string("op_4537"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_115_cast_fp16 = reshape(shape = var_4537, x = linear_174_cast_fp16)[name = string("q_115_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392162112))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393210752))))[name = string("encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393212864)))];
+            tensor<fp16, [1, 56, 1024]> linear_175_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_175_cast_fp16")];
+            tensor<int32, [4]> var_4542 = const()[name = string("op_4542"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_77_cast_fp16 = reshape(shape = var_4542, x = linear_175_cast_fp16)[name = string("k_77_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393214976))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394263616))))[name = string("encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394265728)))];
+            tensor<fp16, [1, 56, 1024]> linear_176_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_176_cast_fp16")];
+            tensor<int32, [4]> var_4547 = const()[name = string("op_4547"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_39_cast_fp16 = reshape(shape = var_4547, x = linear_176_cast_fp16)[name = string("v_39_cast_fp16")];
+            tensor<int32, [4]> value_47_perm_0 = const()[name = string("value_47_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_19_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394267840)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4560_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_u_to_fp16)[name = string("op_4560_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_19_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394269952)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4562_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_v_to_fp16)[name = string("op_4562_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_39_perm_0 = const()[name = string("q_with_bias_v_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_501_transpose_x_0 = const()[name = string("x_501_transpose_x_0"), val = bool(false)];
+            bool x_501_transpose_y_0 = const()[name = string("x_501_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_4564_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394272064))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394385792))))[name = string("op_4564_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_39_cast_fp16 = transpose(perm = q_with_bias_v_39_perm_0, x = var_4562_cast_fp16)[name = string("transpose_191")];
+            tensor<fp16, [1, 8, 14, 111]> x_501_cast_fp16 = matmul(transpose_x = x_501_transpose_x_0, transpose_y = x_501_transpose_y_0, x = q_with_bias_v_39_cast_fp16, y = op_4564_to_fp16_quantized)[name = string("x_501_cast_fp16")];
+            tensor<int32, [8]> x_503_pad_0 = const()[name = string("x_503_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_503_mode_0 = const()[name = string("x_503_mode_0"), val = string("constant")];
+            fp16 const_326_to_fp16 = const()[name = string("const_326_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_503_cast_fp16 = pad(constant_val = const_326_to_fp16, mode = x_503_mode_0, pad = x_503_pad_0, x = x_501_cast_fp16)[name = string("x_503_cast_fp16")];
+            tensor<int32, [4]> var_4572 = const()[name = string("op_4572"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_505_cast_fp16 = reshape(shape = var_4572, x = x_503_cast_fp16)[name = string("x_505_cast_fp16")];
+            tensor<int32, [4]> var_4576_begin_0 = const()[name = string("op_4576_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4576_end_0 = const()[name = string("op_4576_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_4576_end_mask_0 = const()[name = string("op_4576_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_4576_cast_fp16 = slice_by_index(begin = var_4576_begin_0, end = var_4576_end_0, end_mask = var_4576_end_mask_0, x = x_505_cast_fp16)[name = string("op_4576_cast_fp16")];
+            tensor<int32, [4]> var_4577 = const()[name = string("op_4577"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_77_cast_fp16 = reshape(shape = var_4577, x = var_4576_cast_fp16)[name = string("matrix_bd_77_cast_fp16")];
+            bool matrix_ac_39_transpose_x_0 = const()[name = string("matrix_ac_39_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_39_transpose_y_0 = const()[name = string("matrix_ac_39_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_134_perm_0 = const()[name = string("transpose_134_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_135_perm_0 = const()[name = string("transpose_135_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_135 = transpose(perm = transpose_135_perm_0, x = k_77_cast_fp16)[name = string("transpose_189")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_134 = transpose(perm = transpose_134_perm_0, x = var_4560_cast_fp16)[name = string("transpose_190")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_39_cast_fp16 = matmul(transpose_x = matrix_ac_39_transpose_x_0, transpose_y = matrix_ac_39_transpose_y_0, x = transpose_134, y = transpose_135)[name = string("matrix_ac_39_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_79_begin_0 = const()[name = string("matrix_bd_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_79_end_0 = const()[name = string("matrix_bd_79_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_79_end_mask_0 = const()[name = string("matrix_bd_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_79_cast_fp16 = slice_by_index(begin = matrix_bd_79_begin_0, end = matrix_bd_79_end_0, end_mask = matrix_bd_79_end_mask_0, x = matrix_bd_77_cast_fp16)[name = string("matrix_bd_79_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_4586_cast_fp16 = add(x = matrix_ac_39_cast_fp16, y = matrix_bd_79_cast_fp16)[name = string("op_4586_cast_fp16")];
+            fp16 _inversed_scores_77_y_0_to_fp16 = const()[name = string("_inversed_scores_77_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_77_cast_fp16 = mul(x = var_4586_cast_fp16, y = _inversed_scores_77_y_0_to_fp16)[name = string("_inversed_scores_77_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_79_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_77_cast_fp16, cond = mask_11)[name = string("scores_79_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_4592_cast_fp16 = softmax(axis = var_60, x = scores_79_cast_fp16)[name = string("op_4592_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_1029_cast_fp16 = select(a = var_45_to_fp16, b = var_4592_cast_fp16, cond = mask_11)[name = string("input_1029_cast_fp16")];
+            bool x_507_transpose_x_0 = const()[name = string("x_507_transpose_x_0"), val = bool(false)];
+            bool x_507_transpose_y_0 = const()[name = string("x_507_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = v_39_cast_fp16)[name = string("transpose_188")];
+            tensor<fp16, [1, 8, 14, 128]> x_507_cast_fp16 = matmul(transpose_x = x_507_transpose_x_0, transpose_y = x_507_transpose_y_0, x = input_1029_cast_fp16, y = value_47_cast_fp16)[name = string("x_507_cast_fp16")];
+            tensor<int32, [4]> var_4596_perm_0 = const()[name = string("op_4596_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4597 = const()[name = string("op_4597"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_4596_cast_fp16 = transpose(perm = var_4596_perm_0, x = x_507_cast_fp16)[name = string("transpose_187")];
+            tensor<fp16, [1, 14, 1024]> input_1031_cast_fp16 = reshape(shape = var_4597, x = var_4596_cast_fp16)[name = string("input_1031_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394386112))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395434752))))[name = string("encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395436864)))];
+            tensor<fp16, [1, 14, 1024]> linear_178_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized, x = input_1031_cast_fp16)[name = string("linear_178_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1035_cast_fp16 = add(x = input_1025_cast_fp16, y = linear_178_cast_fp16)[name = string("input_1035_cast_fp16")];
+            tensor<int32, [1]> x_511_axes_0 = const()[name = string("x_511_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395438976)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395441088)))];
+            tensor<fp16, [1, 14, 1024]> x_511_cast_fp16 = layer_norm(axes = x_511_axes_0, beta = encoder_layers_19_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_conv_weight_to_fp16, x = input_1035_cast_fp16)[name = string("x_511_cast_fp16")];
+            tensor<int32, [3]> input_1037_perm_0 = const()[name = string("input_1037_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1039_pad_type_0 = const()[name = string("input_1039_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1039_strides_0 = const()[name = string("input_1039_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1039_pad_0 = const()[name = string("input_1039_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1039_dilations_0 = const()[name = string("input_1039_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1039_groups_0 = const()[name = string("input_1039_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395443200))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397540416))))[name = string("encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_1037_cast_fp16 = transpose(perm = input_1037_perm_0, x = x_511_cast_fp16)[name = string("transpose_186")];
+            tensor<fp16, [1, 2048, 14]> input_1039_cast_fp16 = conv(dilations = input_1039_dilations_0, groups = input_1039_groups_0, pad = input_1039_pad_0, pad_type = input_1039_pad_type_0, strides = input_1039_strides_0, weight = encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1037_cast_fp16)[name = string("input_1039_cast_fp16")];
+            int32 x_513_split_num_splits_0 = const()[name = string("x_513_split_num_splits_0"), val = int32(2)];
+            int32 x_513_split_axis_0 = const()[name = string("x_513_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_513_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_513_split_cast_fp16_1 = split(axis = x_513_split_axis_0, num_splits = x_513_split_num_splits_0, x = input_1039_cast_fp16)[name = string("x_513_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_513_split_1_sigmoid_cast_fp16 = sigmoid(x = x_513_split_cast_fp16_1)[name = string("x_513_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_513_cast_fp16 = mul(x = x_513_split_cast_fp16_0, y = x_513_split_1_sigmoid_cast_fp16)[name = string("x_513_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_1041_cast_fp16 = select(a = var_45_to_fp16, b = x_513_cast_fp16, cond = var_576)[name = string("input_1041_cast_fp16")];
+            bool new_x_79_interleave_0 = const()[name = string("new_x_79_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_79_cast_fp16 = concat(axis = var_60, interleave = new_x_79_interleave_0, values = (cache_79_cast_fp16, input_1041_cast_fp16))[name = string("new_x_79_cast_fp16")];
+            tensor<int32, [3]> var_4636_begin_0 = const()[name = string("op_4636_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_4636_end_0 = const()[name = string("op_4636_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_4636_end_mask_0 = const()[name = string("op_4636_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4636_cast_fp16 = slice_by_index(begin = var_4636_begin_0, end = var_4636_end_0, end_mask = var_4636_end_mask_0, x = new_x_79_cast_fp16)[name = string("op_4636_cast_fp16")];
+            string x_515_pad_type_0 = const()[name = string("x_515_pad_type_0"), val = string("valid")];
+            int32 x_515_groups_0 = const()[name = string("x_515_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_515_strides_0 = const()[name = string("x_515_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_515_pad_0 = const()[name = string("x_515_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_515_dilations_0 = const()[name = string("x_515_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397544576))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397553856))))[name = string("encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_515_cast_fp16 = conv(dilations = x_515_dilations_0, groups = x_515_groups_0, pad = x_515_pad_0, pad_type = x_515_pad_type_0, strides = x_515_strides_0, weight = encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_79_cast_fp16)[name = string("x_515_cast_fp16")];
+            tensor<int32, [3]> input_1043_perm_0 = const()[name = string("input_1043_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_517_axes_0 = const()[name = string("x_517_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397555968)))];
+            tensor<fp16, [1024]> encoder_layers_19_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397558080)))];
+            tensor<fp16, [1, 14, 1024]> input_1043_cast_fp16 = transpose(perm = input_1043_perm_0, x = x_515_cast_fp16)[name = string("transpose_185")];
+            tensor<fp16, [1, 14, 1024]> x_517_cast_fp16 = layer_norm(axes = x_517_axes_0, beta = encoder_layers_19_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_conv_batch_norm_weight_to_fp16, x = input_1043_cast_fp16)[name = string("x_517_cast_fp16")];
+            tensor<int32, [3]> input_1045_perm_0 = const()[name = string("input_1045_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_1045_cast_fp16 = transpose(perm = input_1045_perm_0, x = x_517_cast_fp16)[name = string("transpose_184")];
+            tensor<fp16, [1, 1024, 14]> input_1047_cast_fp16 = silu(x = input_1045_cast_fp16)[name = string("input_1047_cast_fp16")];
+            string x_519_pad_type_0 = const()[name = string("x_519_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_519_strides_0 = const()[name = string("x_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_519_pad_0 = const()[name = string("x_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_519_dilations_0 = const()[name = string("x_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_519_groups_0 = const()[name = string("x_519_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397560192))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398608832))))[name = string("encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_519_cast_fp16 = conv(dilations = x_519_dilations_0, groups = x_519_groups_0, pad = x_519_pad_0, pad_type = x_519_pad_type_0, strides = x_519_strides_0, weight = encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1047_cast_fp16)[name = string("x_519_cast_fp16")];
+            tensor<int32, [3]> input_1049_perm_0 = const()[name = string("input_1049_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_1049_cast_fp16 = transpose(perm = input_1049_perm_0, x = x_519_cast_fp16)[name = string("transpose_183")];
+            tensor<fp16, [1, 14, 1024]> input_1051_cast_fp16 = add(x = input_1035_cast_fp16, y = input_1049_cast_fp16)[name = string("input_1051_cast_fp16")];
+            tensor<int32, [1]> input_1053_axes_0 = const()[name = string("input_1053_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398610944)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398613056)))];
+            tensor<fp16, [1, 14, 1024]> input_1053_cast_fp16 = layer_norm(axes = input_1053_axes_0, beta = encoder_layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_feed_forward2_weight_to_fp16, x = input_1051_cast_fp16)[name = string("input_1053_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398615168))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402809536))))[name = string("encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_19_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402817792)))];
+            tensor<fp16, [1, 14, 4096]> linear_179_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1053_cast_fp16)[name = string("linear_179_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1057_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("input_1057_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402826048))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407020416))))[name = string("encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407022528)))];
+            tensor<fp16, [1, 14, 1024]> linear_180_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1057_cast_fp16)[name = string("linear_180_cast_fp16")];
+            fp16 var_4679_to_fp16 = const()[name = string("op_4679_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4680_cast_fp16 = mul(x = linear_180_cast_fp16, y = var_4679_to_fp16)[name = string("op_4680_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1063_cast_fp16 = add(x = input_1051_cast_fp16, y = var_4680_cast_fp16)[name = string("input_1063_cast_fp16")];
+            tensor<int32, [1]> input_1065_axes_0 = const()[name = string("input_1065_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407024640)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407026752)))];
+            tensor<fp16, [1, 14, 1024]> input_1065_cast_fp16 = layer_norm(axes = input_1065_axes_0, beta = encoder_layers_19_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_19_norm_out_weight_to_fp16, x = input_1063_cast_fp16)[name = string("input_1065_cast_fp16")];
+            tensor<int32, [4]> cache_81_begin_0 = const()[name = string("cache_81_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
+            tensor<int32, [4]> cache_81_end_0 = const()[name = string("cache_81_end_0"), val = tensor<int32, [4]>([21, 1, 42, 1024])];
+            tensor<bool, [4]> cache_81_end_mask_0 = const()[name = string("cache_81_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_81_squeeze_mask_0 = const()[name = string("cache_81_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_81_cast_fp16 = slice_by_index(begin = cache_81_begin_0, end = cache_81_end_0, end_mask = cache_81_end_mask_0, squeeze_mask = cache_81_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_81_cast_fp16")];
+            tensor<int32, [4]> cache_83_begin_0 = const()[name = string("cache_83_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
+            tensor<int32, [4]> cache_83_end_0 = const()[name = string("cache_83_end_0"), val = tensor<int32, [4]>([21, 1, 1024, 8])];
+            tensor<bool, [4]> cache_83_end_mask_0 = const()[name = string("cache_83_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_83_squeeze_mask_0 = const()[name = string("cache_83_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_83_cast_fp16 = slice_by_index(begin = cache_83_begin_0, end = cache_83_end_0, end_mask = cache_83_end_mask_0, squeeze_mask = cache_83_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_83_cast_fp16")];
+            tensor<int32, [1]> input_1067_axes_0 = const()[name = string("input_1067_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407028864)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407030976)))];
+            tensor<fp16, [1, 14, 1024]> input_1067_cast_fp16 = layer_norm(axes = input_1067_axes_0, beta = encoder_layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_feed_forward1_weight_to_fp16, x = input_1065_cast_fp16)[name = string("input_1067_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407033088))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411227456))))[name = string("encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_20_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411235712)))];
+            tensor<fp16, [1, 14, 4096]> linear_181_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1067_cast_fp16)[name = string("linear_181_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1071_cast_fp16 = silu(x = linear_181_cast_fp16)[name = string("input_1071_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411243968))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415438336))))[name = string("encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415440448)))];
+            tensor<fp16, [1, 14, 1024]> linear_182_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1071_cast_fp16)[name = string("linear_182_cast_fp16")];
+            fp16 var_4716_to_fp16 = const()[name = string("op_4716_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4717_cast_fp16 = mul(x = linear_182_cast_fp16, y = var_4716_to_fp16)[name = string("op_4717_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1077_cast_fp16 = add(x = input_1065_cast_fp16, y = var_4717_cast_fp16)[name = string("input_1077_cast_fp16")];
+            tensor<int32, [1]> key_41_axes_0 = const()[name = string("key_41_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415442560)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415444672)))];
+            tensor<fp16, [1, 14, 1024]> key_41_cast_fp16 = layer_norm(axes = key_41_axes_0, beta = encoder_layers_20_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_self_att_weight_to_fp16, x = input_1077_cast_fp16)[name = string("key_41_cast_fp16")];
+            bool input_1079_interleave_0 = const()[name = string("input_1079_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_1079_cast_fp16 = concat(axis = var_69, interleave = input_1079_interleave_0, values = (cache_81_cast_fp16, key_41_cast_fp16))[name = string("input_1079_cast_fp16")];
+            tensor<int32, [3]> var_4739_begin_0 = const()[name = string("op_4739_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_4739_end_0 = const()[name = string("op_4739_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4739_end_mask_0 = const()[name = string("op_4739_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_4739_cast_fp16 = slice_by_index(begin = var_4739_begin_0, end = var_4739_end_0, end_mask = var_4739_end_mask_0, x = cache_81_cast_fp16)[name = string("op_4739_cast_fp16")];
+            bool var_4745_interleave_0 = const()[name = string("op_4745_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4745_cast_fp16 = concat(axis = var_69, interleave = var_4745_interleave_0, values = (var_4739_cast_fp16, key_41_cast_fp16))[name = string("op_4745_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415446784))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416495424))))[name = string("encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416497536)))];
+            tensor<fp16, [1, 14, 1024]> linear_183_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized, x = key_41_cast_fp16)[name = string("linear_183_cast_fp16")];
+            tensor<int32, [4]> var_4750 = const()[name = string("op_4750"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_121_cast_fp16 = reshape(shape = var_4750, x = linear_183_cast_fp16)[name = string("q_121_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416499648))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417548288))))[name = string("encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417550400)))];
+            tensor<fp16, [1, 56, 1024]> linear_184_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_184_cast_fp16")];
+            tensor<int32, [4]> var_4755 = const()[name = string("op_4755"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_81_cast_fp16 = reshape(shape = var_4755, x = linear_184_cast_fp16)[name = string("k_81_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417552512))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418601152))))[name = string("encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418603264)))];
+            tensor<fp16, [1, 56, 1024]> linear_185_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_185_cast_fp16")];
+            tensor<int32, [4]> var_4760 = const()[name = string("op_4760"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_41_cast_fp16 = reshape(shape = var_4760, x = linear_185_cast_fp16)[name = string("v_41_cast_fp16")];
+            tensor<int32, [4]> value_49_perm_0 = const()[name = string("value_49_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_20_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418605376)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4773_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_u_to_fp16)[name = string("op_4773_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_20_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418607488)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4775_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_v_to_fp16)[name = string("op_4775_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_41_perm_0 = const()[name = string("q_with_bias_v_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_527_transpose_x_0 = const()[name = string("x_527_transpose_x_0"), val = bool(false)];
+            bool x_527_transpose_y_0 = const()[name = string("x_527_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_4777_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418609600))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418723328))))[name = string("op_4777_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_41_cast_fp16 = transpose(perm = q_with_bias_v_41_perm_0, x = var_4775_cast_fp16)[name = string("transpose_182")];
+            tensor<fp16, [1, 8, 14, 111]> x_527_cast_fp16 = matmul(transpose_x = x_527_transpose_x_0, transpose_y = x_527_transpose_y_0, x = q_with_bias_v_41_cast_fp16, y = op_4777_to_fp16_quantized)[name = string("x_527_cast_fp16")];
+            tensor<int32, [8]> x_529_pad_0 = const()[name = string("x_529_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_529_mode_0 = const()[name = string("x_529_mode_0"), val = string("constant")];
+            fp16 const_339_to_fp16 = const()[name = string("const_339_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_529_cast_fp16 = pad(constant_val = const_339_to_fp16, mode = x_529_mode_0, pad = x_529_pad_0, x = x_527_cast_fp16)[name = string("x_529_cast_fp16")];
+            tensor<int32, [4]> var_4785 = const()[name = string("op_4785"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_531_cast_fp16 = reshape(shape = var_4785, x = x_529_cast_fp16)[name = string("x_531_cast_fp16")];
+            tensor<int32, [4]> var_4789_begin_0 = const()[name = string("op_4789_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4789_end_0 = const()[name = string("op_4789_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_4789_end_mask_0 = const()[name = string("op_4789_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_4789_cast_fp16 = slice_by_index(begin = var_4789_begin_0, end = var_4789_end_0, end_mask = var_4789_end_mask_0, x = x_531_cast_fp16)[name = string("op_4789_cast_fp16")];
+            tensor<int32, [4]> var_4790 = const()[name = string("op_4790"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_81_cast_fp16 = reshape(shape = var_4790, x = var_4789_cast_fp16)[name = string("matrix_bd_81_cast_fp16")];
+            bool matrix_ac_41_transpose_x_0 = const()[name = string("matrix_ac_41_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_41_transpose_y_0 = const()[name = string("matrix_ac_41_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_136_perm_0 = const()[name = string("transpose_136_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_137_perm_0 = const()[name = string("transpose_137_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_137 = transpose(perm = transpose_137_perm_0, x = k_81_cast_fp16)[name = string("transpose_180")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_136 = transpose(perm = transpose_136_perm_0, x = var_4773_cast_fp16)[name = string("transpose_181")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_41_cast_fp16 = matmul(transpose_x = matrix_ac_41_transpose_x_0, transpose_y = matrix_ac_41_transpose_y_0, x = transpose_136, y = transpose_137)[name = string("matrix_ac_41_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_83_begin_0 = const()[name = string("matrix_bd_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_83_end_0 = const()[name = string("matrix_bd_83_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_83_end_mask_0 = const()[name = string("matrix_bd_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_83_cast_fp16 = slice_by_index(begin = matrix_bd_83_begin_0, end = matrix_bd_83_end_0, end_mask = matrix_bd_83_end_mask_0, x = matrix_bd_81_cast_fp16)[name = string("matrix_bd_83_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_4799_cast_fp16 = add(x = matrix_ac_41_cast_fp16, y = matrix_bd_83_cast_fp16)[name = string("op_4799_cast_fp16")];
+            fp16 _inversed_scores_81_y_0_to_fp16 = const()[name = string("_inversed_scores_81_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_81_cast_fp16 = mul(x = var_4799_cast_fp16, y = _inversed_scores_81_y_0_to_fp16)[name = string("_inversed_scores_81_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_83_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_81_cast_fp16, cond = mask_11)[name = string("scores_83_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_4805_cast_fp16 = softmax(axis = var_60, x = scores_83_cast_fp16)[name = string("op_4805_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_1081_cast_fp16 = select(a = var_45_to_fp16, b = var_4805_cast_fp16, cond = mask_11)[name = string("input_1081_cast_fp16")];
+            bool x_533_transpose_x_0 = const()[name = string("x_533_transpose_x_0"), val = bool(false)];
+            bool x_533_transpose_y_0 = const()[name = string("x_533_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = v_41_cast_fp16)[name = string("transpose_179")];
+            tensor<fp16, [1, 8, 14, 128]> x_533_cast_fp16 = matmul(transpose_x = x_533_transpose_x_0, transpose_y = x_533_transpose_y_0, x = input_1081_cast_fp16, y = value_49_cast_fp16)[name = string("x_533_cast_fp16")];
+            tensor<int32, [4]> var_4809_perm_0 = const()[name = string("op_4809_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4810 = const()[name = string("op_4810"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_4809_cast_fp16 = transpose(perm = var_4809_perm_0, x = x_533_cast_fp16)[name = string("transpose_178")];
+            tensor<fp16, [1, 14, 1024]> input_1083_cast_fp16 = reshape(shape = var_4810, x = var_4809_cast_fp16)[name = string("input_1083_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418723648))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419772288))))[name = string("encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419774400)))];
+            tensor<fp16, [1, 14, 1024]> linear_187_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized, x = input_1083_cast_fp16)[name = string("linear_187_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1087_cast_fp16 = add(x = input_1077_cast_fp16, y = linear_187_cast_fp16)[name = string("input_1087_cast_fp16")];
+            tensor<int32, [1]> x_537_axes_0 = const()[name = string("x_537_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419776512)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419778624)))];
+            tensor<fp16, [1, 14, 1024]> x_537_cast_fp16 = layer_norm(axes = x_537_axes_0, beta = encoder_layers_20_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_conv_weight_to_fp16, x = input_1087_cast_fp16)[name = string("x_537_cast_fp16")];
+            tensor<int32, [3]> input_1089_perm_0 = const()[name = string("input_1089_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1091_pad_type_0 = const()[name = string("input_1091_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1091_strides_0 = const()[name = string("input_1091_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1091_pad_0 = const()[name = string("input_1091_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1091_dilations_0 = const()[name = string("input_1091_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1091_groups_0 = const()[name = string("input_1091_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419780736))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421877952))))[name = string("encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_1089_cast_fp16 = transpose(perm = input_1089_perm_0, x = x_537_cast_fp16)[name = string("transpose_177")];
+            tensor<fp16, [1, 2048, 14]> input_1091_cast_fp16 = conv(dilations = input_1091_dilations_0, groups = input_1091_groups_0, pad = input_1091_pad_0, pad_type = input_1091_pad_type_0, strides = input_1091_strides_0, weight = encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1089_cast_fp16)[name = string("input_1091_cast_fp16")];
+            int32 x_539_split_num_splits_0 = const()[name = string("x_539_split_num_splits_0"), val = int32(2)];
+            int32 x_539_split_axis_0 = const()[name = string("x_539_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_539_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_539_split_cast_fp16_1 = split(axis = x_539_split_axis_0, num_splits = x_539_split_num_splits_0, x = input_1091_cast_fp16)[name = string("x_539_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_539_split_1_sigmoid_cast_fp16 = sigmoid(x = x_539_split_cast_fp16_1)[name = string("x_539_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_539_cast_fp16 = mul(x = x_539_split_cast_fp16_0, y = x_539_split_1_sigmoid_cast_fp16)[name = string("x_539_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_1093_cast_fp16 = select(a = var_45_to_fp16, b = x_539_cast_fp16, cond = var_576)[name = string("input_1093_cast_fp16")];
+            bool new_x_83_interleave_0 = const()[name = string("new_x_83_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_83_cast_fp16 = concat(axis = var_60, interleave = new_x_83_interleave_0, values = (cache_83_cast_fp16, input_1093_cast_fp16))[name = string("new_x_83_cast_fp16")];
+            tensor<int32, [3]> var_4849_begin_0 = const()[name = string("op_4849_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_4849_end_0 = const()[name = string("op_4849_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_4849_end_mask_0 = const()[name = string("op_4849_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4849_cast_fp16 = slice_by_index(begin = var_4849_begin_0, end = var_4849_end_0, end_mask = var_4849_end_mask_0, x = new_x_83_cast_fp16)[name = string("op_4849_cast_fp16")];
+            string x_541_pad_type_0 = const()[name = string("x_541_pad_type_0"), val = string("valid")];
+            int32 x_541_groups_0 = const()[name = string("x_541_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_541_strides_0 = const()[name = string("x_541_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_541_pad_0 = const()[name = string("x_541_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_541_dilations_0 = const()[name = string("x_541_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421882112))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421891392))))[name = string("encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_541_cast_fp16 = conv(dilations = x_541_dilations_0, groups = x_541_groups_0, pad = x_541_pad_0, pad_type = x_541_pad_type_0, strides = x_541_strides_0, weight = encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_83_cast_fp16)[name = string("x_541_cast_fp16")];
+            tensor<int32, [3]> input_1095_perm_0 = const()[name = string("input_1095_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_543_axes_0 = const()[name = string("x_543_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421893504)))];
+            tensor<fp16, [1024]> encoder_layers_20_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421895616)))];
+            tensor<fp16, [1, 14, 1024]> input_1095_cast_fp16 = transpose(perm = input_1095_perm_0, x = x_541_cast_fp16)[name = string("transpose_176")];
+            tensor<fp16, [1, 14, 1024]> x_543_cast_fp16 = layer_norm(axes = x_543_axes_0, beta = encoder_layers_20_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_conv_batch_norm_weight_to_fp16, x = input_1095_cast_fp16)[name = string("x_543_cast_fp16")];
+            tensor<int32, [3]> input_1097_perm_0 = const()[name = string("input_1097_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_1097_cast_fp16 = transpose(perm = input_1097_perm_0, x = x_543_cast_fp16)[name = string("transpose_175")];
+            tensor<fp16, [1, 1024, 14]> input_1099_cast_fp16 = silu(x = input_1097_cast_fp16)[name = string("input_1099_cast_fp16")];
+            string x_545_pad_type_0 = const()[name = string("x_545_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_545_strides_0 = const()[name = string("x_545_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_545_pad_0 = const()[name = string("x_545_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_545_dilations_0 = const()[name = string("x_545_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_545_groups_0 = const()[name = string("x_545_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421897728))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422946368))))[name = string("encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_545_cast_fp16 = conv(dilations = x_545_dilations_0, groups = x_545_groups_0, pad = x_545_pad_0, pad_type = x_545_pad_type_0, strides = x_545_strides_0, weight = encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1099_cast_fp16)[name = string("x_545_cast_fp16")];
+            tensor<int32, [3]> input_1101_perm_0 = const()[name = string("input_1101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_1101_cast_fp16 = transpose(perm = input_1101_perm_0, x = x_545_cast_fp16)[name = string("transpose_174")];
+            tensor<fp16, [1, 14, 1024]> input_1103_cast_fp16 = add(x = input_1087_cast_fp16, y = input_1101_cast_fp16)[name = string("input_1103_cast_fp16")];
+            tensor<int32, [1]> input_1105_axes_0 = const()[name = string("input_1105_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422948480)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422950592)))];
+            tensor<fp16, [1, 14, 1024]> input_1105_cast_fp16 = layer_norm(axes = input_1105_axes_0, beta = encoder_layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_feed_forward2_weight_to_fp16, x = input_1103_cast_fp16)[name = string("input_1105_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422952704))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427147072))))[name = string("encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_20_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427155328)))];
+            tensor<fp16, [1, 14, 4096]> linear_188_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1105_cast_fp16)[name = string("linear_188_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1109_cast_fp16 = silu(x = linear_188_cast_fp16)[name = string("input_1109_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427163584))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431357952))))[name = string("encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431360064)))];
+            tensor<fp16, [1, 14, 1024]> linear_189_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1109_cast_fp16)[name = string("linear_189_cast_fp16")];
+            fp16 var_4892_to_fp16 = const()[name = string("op_4892_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4893_cast_fp16 = mul(x = linear_189_cast_fp16, y = var_4892_to_fp16)[name = string("op_4893_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1115_cast_fp16 = add(x = input_1103_cast_fp16, y = var_4893_cast_fp16)[name = string("input_1115_cast_fp16")];
+            tensor<int32, [1]> input_1117_axes_0 = const()[name = string("input_1117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431362176)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431364288)))];
+            tensor<fp16, [1, 14, 1024]> input_1117_cast_fp16 = layer_norm(axes = input_1117_axes_0, beta = encoder_layers_20_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_20_norm_out_weight_to_fp16, x = input_1115_cast_fp16)[name = string("input_1117_cast_fp16")];
+            tensor<int32, [4]> cache_85_begin_0 = const()[name = string("cache_85_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
+            tensor<int32, [4]> cache_85_end_0 = const()[name = string("cache_85_end_0"), val = tensor<int32, [4]>([22, 1, 42, 1024])];
+            tensor<bool, [4]> cache_85_end_mask_0 = const()[name = string("cache_85_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_85_squeeze_mask_0 = const()[name = string("cache_85_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_85_cast_fp16 = slice_by_index(begin = cache_85_begin_0, end = cache_85_end_0, end_mask = cache_85_end_mask_0, squeeze_mask = cache_85_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_85_cast_fp16")];
+            tensor<int32, [4]> cache_87_begin_0 = const()[name = string("cache_87_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
+            tensor<int32, [4]> cache_87_end_0 = const()[name = string("cache_87_end_0"), val = tensor<int32, [4]>([22, 1, 1024, 8])];
+            tensor<bool, [4]> cache_87_end_mask_0 = const()[name = string("cache_87_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_87_squeeze_mask_0 = const()[name = string("cache_87_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_87_cast_fp16 = slice_by_index(begin = cache_87_begin_0, end = cache_87_end_0, end_mask = cache_87_end_mask_0, squeeze_mask = cache_87_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_87_cast_fp16")];
+            tensor<int32, [1]> input_1119_axes_0 = const()[name = string("input_1119_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431366400)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431368512)))];
+            tensor<fp16, [1, 14, 1024]> input_1119_cast_fp16 = layer_norm(axes = input_1119_axes_0, beta = encoder_layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_feed_forward1_weight_to_fp16, x = input_1117_cast_fp16)[name = string("input_1119_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431370624))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435564992))))[name = string("encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_21_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435573248)))];
+            tensor<fp16, [1, 14, 4096]> linear_190_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1119_cast_fp16)[name = string("linear_190_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1123_cast_fp16 = silu(x = linear_190_cast_fp16)[name = string("input_1123_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435581504))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439775872))))[name = string("encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439777984)))];
+            tensor<fp16, [1, 14, 1024]> linear_191_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1123_cast_fp16)[name = string("linear_191_cast_fp16")];
+            fp16 var_4929_to_fp16 = const()[name = string("op_4929_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_4930_cast_fp16 = mul(x = linear_191_cast_fp16, y = var_4929_to_fp16)[name = string("op_4930_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1129_cast_fp16 = add(x = input_1117_cast_fp16, y = var_4930_cast_fp16)[name = string("input_1129_cast_fp16")];
+            tensor<int32, [1]> key_43_axes_0 = const()[name = string("key_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439780096)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439782208)))];
+            tensor<fp16, [1, 14, 1024]> key_43_cast_fp16 = layer_norm(axes = key_43_axes_0, beta = encoder_layers_21_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_self_att_weight_to_fp16, x = input_1129_cast_fp16)[name = string("key_43_cast_fp16")];
+            bool input_1131_interleave_0 = const()[name = string("input_1131_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_1131_cast_fp16 = concat(axis = var_69, interleave = input_1131_interleave_0, values = (cache_85_cast_fp16, key_43_cast_fp16))[name = string("input_1131_cast_fp16")];
+            tensor<int32, [3]> var_4952_begin_0 = const()[name = string("op_4952_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_4952_end_0 = const()[name = string("op_4952_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4952_end_mask_0 = const()[name = string("op_4952_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_4952_cast_fp16 = slice_by_index(begin = var_4952_begin_0, end = var_4952_end_0, end_mask = var_4952_end_mask_0, x = cache_85_cast_fp16)[name = string("op_4952_cast_fp16")];
+            bool var_4958_interleave_0 = const()[name = string("op_4958_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4958_cast_fp16 = concat(axis = var_69, interleave = var_4958_interleave_0, values = (var_4952_cast_fp16, key_43_cast_fp16))[name = string("op_4958_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439784320))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440832960))))[name = string("encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440835072)))];
+            tensor<fp16, [1, 14, 1024]> linear_192_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized, x = key_43_cast_fp16)[name = string("linear_192_cast_fp16")];
+            tensor<int32, [4]> var_4963 = const()[name = string("op_4963"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_127_cast_fp16 = reshape(shape = var_4963, x = linear_192_cast_fp16)[name = string("q_127_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440837184))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441885824))))[name = string("encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441887936)))];
+            tensor<fp16, [1, 56, 1024]> linear_193_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_193_cast_fp16")];
+            tensor<int32, [4]> var_4968 = const()[name = string("op_4968"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_85_cast_fp16 = reshape(shape = var_4968, x = linear_193_cast_fp16)[name = string("k_85_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441890048))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442938688))))[name = string("encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442940800)))];
+            tensor<fp16, [1, 56, 1024]> linear_194_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_194_cast_fp16")];
+            tensor<int32, [4]> var_4973 = const()[name = string("op_4973"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_43_cast_fp16 = reshape(shape = var_4973, x = linear_194_cast_fp16)[name = string("v_43_cast_fp16")];
+            tensor<int32, [4]> value_51_perm_0 = const()[name = string("value_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_21_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442942912)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4986_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_u_to_fp16)[name = string("op_4986_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_21_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442945024)))];
+            tensor<fp16, [1, 14, 8, 128]> var_4988_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_v_to_fp16)[name = string("op_4988_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_43_perm_0 = const()[name = string("q_with_bias_v_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_553_transpose_x_0 = const()[name = string("x_553_transpose_x_0"), val = bool(false)];
+            bool x_553_transpose_y_0 = const()[name = string("x_553_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_4990_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442947136))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443060864))))[name = string("op_4990_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_43_cast_fp16 = transpose(perm = q_with_bias_v_43_perm_0, x = var_4988_cast_fp16)[name = string("transpose_173")];
+            tensor<fp16, [1, 8, 14, 111]> x_553_cast_fp16 = matmul(transpose_x = x_553_transpose_x_0, transpose_y = x_553_transpose_y_0, x = q_with_bias_v_43_cast_fp16, y = op_4990_to_fp16_quantized)[name = string("x_553_cast_fp16")];
+            tensor<int32, [8]> x_555_pad_0 = const()[name = string("x_555_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_555_mode_0 = const()[name = string("x_555_mode_0"), val = string("constant")];
+            fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_555_cast_fp16 = pad(constant_val = const_352_to_fp16, mode = x_555_mode_0, pad = x_555_pad_0, x = x_553_cast_fp16)[name = string("x_555_cast_fp16")];
+            tensor<int32, [4]> var_4998 = const()[name = string("op_4998"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_557_cast_fp16 = reshape(shape = var_4998, x = x_555_cast_fp16)[name = string("x_557_cast_fp16")];
+            tensor<int32, [4]> var_5002_begin_0 = const()[name = string("op_5002_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5002_end_0 = const()[name = string("op_5002_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_5002_end_mask_0 = const()[name = string("op_5002_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_5002_cast_fp16 = slice_by_index(begin = var_5002_begin_0, end = var_5002_end_0, end_mask = var_5002_end_mask_0, x = x_557_cast_fp16)[name = string("op_5002_cast_fp16")];
+            tensor<int32, [4]> var_5003 = const()[name = string("op_5003"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_85_cast_fp16 = reshape(shape = var_5003, x = var_5002_cast_fp16)[name = string("matrix_bd_85_cast_fp16")];
+            bool matrix_ac_43_transpose_x_0 = const()[name = string("matrix_ac_43_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_43_transpose_y_0 = const()[name = string("matrix_ac_43_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_138_perm_0 = const()[name = string("transpose_138_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_139_perm_0 = const()[name = string("transpose_139_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_139 = transpose(perm = transpose_139_perm_0, x = k_85_cast_fp16)[name = string("transpose_171")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_138 = transpose(perm = transpose_138_perm_0, x = var_4986_cast_fp16)[name = string("transpose_172")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_43_cast_fp16 = matmul(transpose_x = matrix_ac_43_transpose_x_0, transpose_y = matrix_ac_43_transpose_y_0, x = transpose_138, y = transpose_139)[name = string("matrix_ac_43_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_87_begin_0 = const()[name = string("matrix_bd_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_87_end_0 = const()[name = string("matrix_bd_87_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_87_end_mask_0 = const()[name = string("matrix_bd_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_87_cast_fp16 = slice_by_index(begin = matrix_bd_87_begin_0, end = matrix_bd_87_end_0, end_mask = matrix_bd_87_end_mask_0, x = matrix_bd_85_cast_fp16)[name = string("matrix_bd_87_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_5012_cast_fp16 = add(x = matrix_ac_43_cast_fp16, y = matrix_bd_87_cast_fp16)[name = string("op_5012_cast_fp16")];
+            fp16 _inversed_scores_85_y_0_to_fp16 = const()[name = string("_inversed_scores_85_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_85_cast_fp16 = mul(x = var_5012_cast_fp16, y = _inversed_scores_85_y_0_to_fp16)[name = string("_inversed_scores_85_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_87_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_85_cast_fp16, cond = mask_11)[name = string("scores_87_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_5018_cast_fp16 = softmax(axis = var_60, x = scores_87_cast_fp16)[name = string("op_5018_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_1133_cast_fp16 = select(a = var_45_to_fp16, b = var_5018_cast_fp16, cond = mask_11)[name = string("input_1133_cast_fp16")];
+            bool x_559_transpose_x_0 = const()[name = string("x_559_transpose_x_0"), val = bool(false)];
+            bool x_559_transpose_y_0 = const()[name = string("x_559_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = v_43_cast_fp16)[name = string("transpose_170")];
+            tensor<fp16, [1, 8, 14, 128]> x_559_cast_fp16 = matmul(transpose_x = x_559_transpose_x_0, transpose_y = x_559_transpose_y_0, x = input_1133_cast_fp16, y = value_51_cast_fp16)[name = string("x_559_cast_fp16")];
+            tensor<int32, [4]> var_5022_perm_0 = const()[name = string("op_5022_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5023 = const()[name = string("op_5023"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_5022_cast_fp16 = transpose(perm = var_5022_perm_0, x = x_559_cast_fp16)[name = string("transpose_169")];
+            tensor<fp16, [1, 14, 1024]> input_1135_cast_fp16 = reshape(shape = var_5023, x = var_5022_cast_fp16)[name = string("input_1135_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443061184)))];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445158400)))];
+            tensor<fp16, [1, 14, 1024]> linear_196_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_out_weight_to_fp16, x = input_1135_cast_fp16)[name = string("linear_196_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1139_cast_fp16 = add(x = input_1129_cast_fp16, y = linear_196_cast_fp16)[name = string("input_1139_cast_fp16")];
+            tensor<int32, [1]> x_563_axes_0 = const()[name = string("x_563_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445160512)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445162624)))];
+            tensor<fp16, [1, 14, 1024]> x_563_cast_fp16 = layer_norm(axes = x_563_axes_0, beta = encoder_layers_21_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_conv_weight_to_fp16, x = input_1139_cast_fp16)[name = string("x_563_cast_fp16")];
+            tensor<int32, [3]> input_1141_perm_0 = const()[name = string("input_1141_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1143_pad_type_0 = const()[name = string("input_1143_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1143_strides_0 = const()[name = string("input_1143_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1143_pad_0 = const()[name = string("input_1143_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1143_dilations_0 = const()[name = string("input_1143_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1143_groups_0 = const()[name = string("input_1143_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445164736))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447261952))))[name = string("encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_1141_cast_fp16 = transpose(perm = input_1141_perm_0, x = x_563_cast_fp16)[name = string("transpose_168")];
+            tensor<fp16, [1, 2048, 14]> input_1143_cast_fp16 = conv(dilations = input_1143_dilations_0, groups = input_1143_groups_0, pad = input_1143_pad_0, pad_type = input_1143_pad_type_0, strides = input_1143_strides_0, weight = encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1141_cast_fp16)[name = string("input_1143_cast_fp16")];
+            int32 x_565_split_num_splits_0 = const()[name = string("x_565_split_num_splits_0"), val = int32(2)];
+            int32 x_565_split_axis_0 = const()[name = string("x_565_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_565_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_565_split_cast_fp16_1 = split(axis = x_565_split_axis_0, num_splits = x_565_split_num_splits_0, x = input_1143_cast_fp16)[name = string("x_565_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_565_split_1_sigmoid_cast_fp16 = sigmoid(x = x_565_split_cast_fp16_1)[name = string("x_565_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_565_cast_fp16 = mul(x = x_565_split_cast_fp16_0, y = x_565_split_1_sigmoid_cast_fp16)[name = string("x_565_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_1145_cast_fp16 = select(a = var_45_to_fp16, b = x_565_cast_fp16, cond = var_576)[name = string("input_1145_cast_fp16")];
+            bool new_x_87_interleave_0 = const()[name = string("new_x_87_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_87_cast_fp16 = concat(axis = var_60, interleave = new_x_87_interleave_0, values = (cache_87_cast_fp16, input_1145_cast_fp16))[name = string("new_x_87_cast_fp16")];
+            tensor<int32, [3]> var_5062_begin_0 = const()[name = string("op_5062_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_5062_end_0 = const()[name = string("op_5062_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_5062_end_mask_0 = const()[name = string("op_5062_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_5062_cast_fp16 = slice_by_index(begin = var_5062_begin_0, end = var_5062_end_0, end_mask = var_5062_end_mask_0, x = new_x_87_cast_fp16)[name = string("op_5062_cast_fp16")];
+            string x_567_pad_type_0 = const()[name = string("x_567_pad_type_0"), val = string("valid")];
+            int32 x_567_groups_0 = const()[name = string("x_567_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_567_strides_0 = const()[name = string("x_567_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_567_pad_0 = const()[name = string("x_567_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_567_dilations_0 = const()[name = string("x_567_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447266112))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447275392))))[name = string("encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_567_cast_fp16 = conv(dilations = x_567_dilations_0, groups = x_567_groups_0, pad = x_567_pad_0, pad_type = x_567_pad_type_0, strides = x_567_strides_0, weight = encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_87_cast_fp16)[name = string("x_567_cast_fp16")];
+            tensor<int32, [3]> input_1147_perm_0 = const()[name = string("input_1147_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_569_axes_0 = const()[name = string("x_569_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447277504)))];
+            tensor<fp16, [1024]> encoder_layers_21_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447279616)))];
+            tensor<fp16, [1, 14, 1024]> input_1147_cast_fp16 = transpose(perm = input_1147_perm_0, x = x_567_cast_fp16)[name = string("transpose_167")];
+            tensor<fp16, [1, 14, 1024]> x_569_cast_fp16 = layer_norm(axes = x_569_axes_0, beta = encoder_layers_21_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_conv_batch_norm_weight_to_fp16, x = input_1147_cast_fp16)[name = string("x_569_cast_fp16")];
+            tensor<int32, [3]> input_1149_perm_0 = const()[name = string("input_1149_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_1149_cast_fp16 = transpose(perm = input_1149_perm_0, x = x_569_cast_fp16)[name = string("transpose_166")];
+            tensor<fp16, [1, 1024, 14]> input_1151_cast_fp16 = silu(x = input_1149_cast_fp16)[name = string("input_1151_cast_fp16")];
+            string x_571_pad_type_0 = const()[name = string("x_571_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_571_strides_0 = const()[name = string("x_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_571_pad_0 = const()[name = string("x_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_571_dilations_0 = const()[name = string("x_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_571_groups_0 = const()[name = string("x_571_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447281728))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448330368))))[name = string("encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_571_cast_fp16 = conv(dilations = x_571_dilations_0, groups = x_571_groups_0, pad = x_571_pad_0, pad_type = x_571_pad_type_0, strides = x_571_strides_0, weight = encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1151_cast_fp16)[name = string("x_571_cast_fp16")];
+            tensor<int32, [3]> input_1153_perm_0 = const()[name = string("input_1153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_1153_cast_fp16 = transpose(perm = input_1153_perm_0, x = x_571_cast_fp16)[name = string("transpose_165")];
+            tensor<fp16, [1, 14, 1024]> input_1155_cast_fp16 = add(x = input_1139_cast_fp16, y = input_1153_cast_fp16)[name = string("input_1155_cast_fp16")];
+            tensor<int32, [1]> input_1157_axes_0 = const()[name = string("input_1157_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448332480)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448334592)))];
+            tensor<fp16, [1, 14, 1024]> input_1157_cast_fp16 = layer_norm(axes = input_1157_axes_0, beta = encoder_layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_feed_forward2_weight_to_fp16, x = input_1155_cast_fp16)[name = string("input_1157_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_21_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448336704)))];
+            tensor<fp16, [4096]> encoder_layers_21_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456725376)))];
+            tensor<fp16, [1, 14, 4096]> linear_197_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear1_weight_to_fp16, x = input_1157_cast_fp16)[name = string("linear_197_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1161_cast_fp16 = silu(x = linear_197_cast_fp16)[name = string("input_1161_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_21_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456733632)))];
+            tensor<fp16, [1024]> encoder_layers_21_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465122304)))];
+            tensor<fp16, [1, 14, 1024]> linear_198_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear2_weight_to_fp16, x = input_1161_cast_fp16)[name = string("linear_198_cast_fp16")];
+            fp16 var_5105_to_fp16 = const()[name = string("op_5105_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_5106_cast_fp16 = mul(x = linear_198_cast_fp16, y = var_5105_to_fp16)[name = string("op_5106_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1167_cast_fp16 = add(x = input_1155_cast_fp16, y = var_5106_cast_fp16)[name = string("input_1167_cast_fp16")];
+            tensor<int32, [1]> input_1169_axes_0 = const()[name = string("input_1169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465124416)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465126528)))];
+            tensor<fp16, [1, 14, 1024]> input_1169_cast_fp16 = layer_norm(axes = input_1169_axes_0, beta = encoder_layers_21_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_21_norm_out_weight_to_fp16, x = input_1167_cast_fp16)[name = string("input_1169_cast_fp16")];
+            tensor<int32, [4]> cache_89_begin_0 = const()[name = string("cache_89_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
+            tensor<int32, [4]> cache_89_end_0 = const()[name = string("cache_89_end_0"), val = tensor<int32, [4]>([23, 1, 42, 1024])];
+            tensor<bool, [4]> cache_89_end_mask_0 = const()[name = string("cache_89_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_89_squeeze_mask_0 = const()[name = string("cache_89_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_89_cast_fp16 = slice_by_index(begin = cache_89_begin_0, end = cache_89_end_0, end_mask = cache_89_end_mask_0, squeeze_mask = cache_89_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_89_cast_fp16")];
+            tensor<int32, [4]> cache_91_begin_0 = const()[name = string("cache_91_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
+            tensor<int32, [4]> cache_91_end_0 = const()[name = string("cache_91_end_0"), val = tensor<int32, [4]>([23, 1, 1024, 8])];
+            tensor<bool, [4]> cache_91_end_mask_0 = const()[name = string("cache_91_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_91_squeeze_mask_0 = const()[name = string("cache_91_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_91_cast_fp16 = slice_by_index(begin = cache_91_begin_0, end = cache_91_end_0, end_mask = cache_91_end_mask_0, squeeze_mask = cache_91_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_91_cast_fp16")];
+            tensor<int32, [1]> input_1171_axes_0 = const()[name = string("input_1171_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465128640)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465130752)))];
+            tensor<fp16, [1, 14, 1024]> input_1171_cast_fp16 = layer_norm(axes = input_1171_axes_0, beta = encoder_layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_feed_forward1_weight_to_fp16, x = input_1169_cast_fp16)[name = string("input_1171_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_22_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465132864)))];
+            tensor<fp16, [4096]> encoder_layers_22_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473521536)))];
+            tensor<fp16, [1, 14, 4096]> linear_199_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear1_weight_to_fp16, x = input_1171_cast_fp16)[name = string("linear_199_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1175_cast_fp16 = silu(x = linear_199_cast_fp16)[name = string("input_1175_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_22_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473529792)))];
+            tensor<fp16, [1024]> encoder_layers_22_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481918464)))];
+            tensor<fp16, [1, 14, 1024]> linear_200_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear2_weight_to_fp16, x = input_1175_cast_fp16)[name = string("linear_200_cast_fp16")];
+            fp16 var_5142_to_fp16 = const()[name = string("op_5142_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_5143_cast_fp16 = mul(x = linear_200_cast_fp16, y = var_5142_to_fp16)[name = string("op_5143_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1181_cast_fp16 = add(x = input_1169_cast_fp16, y = var_5143_cast_fp16)[name = string("input_1181_cast_fp16")];
+            tensor<int32, [1]> key_45_axes_0 = const()[name = string("key_45_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481920576)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481922688)))];
+            tensor<fp16, [1, 14, 1024]> key_45_cast_fp16 = layer_norm(axes = key_45_axes_0, beta = encoder_layers_22_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_self_att_weight_to_fp16, x = input_1181_cast_fp16)[name = string("key_45_cast_fp16")];
+            bool input_1183_interleave_0 = const()[name = string("input_1183_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_1183_cast_fp16 = concat(axis = var_69, interleave = input_1183_interleave_0, values = (cache_89_cast_fp16, key_45_cast_fp16))[name = string("input_1183_cast_fp16")];
+            tensor<int32, [3]> var_5165_begin_0 = const()[name = string("op_5165_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_5165_end_0 = const()[name = string("op_5165_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_5165_end_mask_0 = const()[name = string("op_5165_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_5165_cast_fp16 = slice_by_index(begin = var_5165_begin_0, end = var_5165_end_0, end_mask = var_5165_end_mask_0, x = cache_89_cast_fp16)[name = string("op_5165_cast_fp16")];
+            bool var_5171_interleave_0 = const()[name = string("op_5171_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_5171_cast_fp16 = concat(axis = var_69, interleave = var_5171_interleave_0, values = (var_5165_cast_fp16, key_45_cast_fp16))[name = string("op_5171_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481924800)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484022016)))];
+            tensor<fp16, [1, 14, 1024]> linear_201_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_q_weight_to_fp16, x = key_45_cast_fp16)[name = string("linear_201_cast_fp16")];
+            tensor<int32, [4]> var_5176 = const()[name = string("op_5176"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_133_cast_fp16 = reshape(shape = var_5176, x = linear_201_cast_fp16)[name = string("q_133_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484024128)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486121344)))];
+            tensor<fp16, [1, 56, 1024]> linear_202_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_k_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_202_cast_fp16")];
+            tensor<int32, [4]> var_5181 = const()[name = string("op_5181"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_89_cast_fp16 = reshape(shape = var_5181, x = linear_202_cast_fp16)[name = string("k_89_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486123456)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488220672)))];
+            tensor<fp16, [1, 56, 1024]> linear_203_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_v_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_203_cast_fp16")];
+            tensor<int32, [4]> var_5186 = const()[name = string("op_5186"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_45_cast_fp16 = reshape(shape = var_5186, x = linear_203_cast_fp16)[name = string("v_45_cast_fp16")];
+            tensor<int32, [4]> value_53_perm_0 = const()[name = string("value_53_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_22_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488222784)))];
+            tensor<fp16, [1, 14, 8, 128]> var_5199_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_u_to_fp16)[name = string("op_5199_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_22_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488224896)))];
+            tensor<fp16, [1, 14, 8, 128]> var_5201_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_v_to_fp16)[name = string("op_5201_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_45_perm_0 = const()[name = string("q_with_bias_v_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_579_transpose_x_0 = const()[name = string("x_579_transpose_x_0"), val = bool(false)];
+            bool x_579_transpose_y_0 = const()[name = string("x_579_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_5203_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488227008))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488340736))))[name = string("op_5203_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_45_cast_fp16 = transpose(perm = q_with_bias_v_45_perm_0, x = var_5201_cast_fp16)[name = string("transpose_164")];
+            tensor<fp16, [1, 8, 14, 111]> x_579_cast_fp16 = matmul(transpose_x = x_579_transpose_x_0, transpose_y = x_579_transpose_y_0, x = q_with_bias_v_45_cast_fp16, y = op_5203_to_fp16_quantized)[name = string("x_579_cast_fp16")];
+            tensor<int32, [8]> x_581_pad_0 = const()[name = string("x_581_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_581_mode_0 = const()[name = string("x_581_mode_0"), val = string("constant")];
+            fp16 const_365_to_fp16 = const()[name = string("const_365_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_581_cast_fp16 = pad(constant_val = const_365_to_fp16, mode = x_581_mode_0, pad = x_581_pad_0, x = x_579_cast_fp16)[name = string("x_581_cast_fp16")];
+            tensor<int32, [4]> var_5211 = const()[name = string("op_5211"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_583_cast_fp16 = reshape(shape = var_5211, x = x_581_cast_fp16)[name = string("x_583_cast_fp16")];
+            tensor<int32, [4]> var_5215_begin_0 = const()[name = string("op_5215_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5215_end_0 = const()[name = string("op_5215_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_5215_end_mask_0 = const()[name = string("op_5215_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_5215_cast_fp16 = slice_by_index(begin = var_5215_begin_0, end = var_5215_end_0, end_mask = var_5215_end_mask_0, x = x_583_cast_fp16)[name = string("op_5215_cast_fp16")];
+            tensor<int32, [4]> var_5216 = const()[name = string("op_5216"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_89_cast_fp16 = reshape(shape = var_5216, x = var_5215_cast_fp16)[name = string("matrix_bd_89_cast_fp16")];
+            bool matrix_ac_45_transpose_x_0 = const()[name = string("matrix_ac_45_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_45_transpose_y_0 = const()[name = string("matrix_ac_45_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_140_perm_0 = const()[name = string("transpose_140_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_141_perm_0 = const()[name = string("transpose_141_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_141 = transpose(perm = transpose_141_perm_0, x = k_89_cast_fp16)[name = string("transpose_162")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_140 = transpose(perm = transpose_140_perm_0, x = var_5199_cast_fp16)[name = string("transpose_163")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_45_cast_fp16 = matmul(transpose_x = matrix_ac_45_transpose_x_0, transpose_y = matrix_ac_45_transpose_y_0, x = transpose_140, y = transpose_141)[name = string("matrix_ac_45_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_91_begin_0 = const()[name = string("matrix_bd_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_91_end_0 = const()[name = string("matrix_bd_91_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_91_end_mask_0 = const()[name = string("matrix_bd_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_91_cast_fp16 = slice_by_index(begin = matrix_bd_91_begin_0, end = matrix_bd_91_end_0, end_mask = matrix_bd_91_end_mask_0, x = matrix_bd_89_cast_fp16)[name = string("matrix_bd_91_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_5225_cast_fp16 = add(x = matrix_ac_45_cast_fp16, y = matrix_bd_91_cast_fp16)[name = string("op_5225_cast_fp16")];
+            fp16 _inversed_scores_89_y_0_to_fp16 = const()[name = string("_inversed_scores_89_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_89_cast_fp16 = mul(x = var_5225_cast_fp16, y = _inversed_scores_89_y_0_to_fp16)[name = string("_inversed_scores_89_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_91_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_89_cast_fp16, cond = mask_11)[name = string("scores_91_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_5231_cast_fp16 = softmax(axis = var_60, x = scores_91_cast_fp16)[name = string("op_5231_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_1185_cast_fp16 = select(a = var_45_to_fp16, b = var_5231_cast_fp16, cond = mask_11)[name = string("input_1185_cast_fp16")];
+            bool x_585_transpose_x_0 = const()[name = string("x_585_transpose_x_0"), val = bool(false)];
+            bool x_585_transpose_y_0 = const()[name = string("x_585_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = v_45_cast_fp16)[name = string("transpose_161")];
+            tensor<fp16, [1, 8, 14, 128]> x_585_cast_fp16 = matmul(transpose_x = x_585_transpose_x_0, transpose_y = x_585_transpose_y_0, x = input_1185_cast_fp16, y = value_53_cast_fp16)[name = string("x_585_cast_fp16")];
+            tensor<int32, [4]> var_5235_perm_0 = const()[name = string("op_5235_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5236 = const()[name = string("op_5236"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_5235_cast_fp16 = transpose(perm = var_5235_perm_0, x = x_585_cast_fp16)[name = string("transpose_160")];
+            tensor<fp16, [1, 14, 1024]> input_1187_cast_fp16 = reshape(shape = var_5236, x = var_5235_cast_fp16)[name = string("input_1187_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488341056)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490438272)))];
+            tensor<fp16, [1, 14, 1024]> linear_205_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_out_weight_to_fp16, x = input_1187_cast_fp16)[name = string("linear_205_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1191_cast_fp16 = add(x = input_1181_cast_fp16, y = linear_205_cast_fp16)[name = string("input_1191_cast_fp16")];
+            tensor<int32, [1]> x_589_axes_0 = const()[name = string("x_589_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490440384)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490442496)))];
+            tensor<fp16, [1, 14, 1024]> x_589_cast_fp16 = layer_norm(axes = x_589_axes_0, beta = encoder_layers_22_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_conv_weight_to_fp16, x = input_1191_cast_fp16)[name = string("x_589_cast_fp16")];
+            tensor<int32, [3]> input_1193_perm_0 = const()[name = string("input_1193_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1195_pad_type_0 = const()[name = string("input_1195_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1195_strides_0 = const()[name = string("input_1195_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1195_pad_0 = const()[name = string("input_1195_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1195_dilations_0 = const()[name = string("input_1195_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1195_groups_0 = const()[name = string("input_1195_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490444608))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492541824))))[name = string("encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_1193_cast_fp16 = transpose(perm = input_1193_perm_0, x = x_589_cast_fp16)[name = string("transpose_159")];
+            tensor<fp16, [1, 2048, 14]> input_1195_cast_fp16 = conv(dilations = input_1195_dilations_0, groups = input_1195_groups_0, pad = input_1195_pad_0, pad_type = input_1195_pad_type_0, strides = input_1195_strides_0, weight = encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1193_cast_fp16)[name = string("input_1195_cast_fp16")];
+            int32 x_591_split_num_splits_0 = const()[name = string("x_591_split_num_splits_0"), val = int32(2)];
+            int32 x_591_split_axis_0 = const()[name = string("x_591_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_591_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_591_split_cast_fp16_1 = split(axis = x_591_split_axis_0, num_splits = x_591_split_num_splits_0, x = input_1195_cast_fp16)[name = string("x_591_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_591_split_1_sigmoid_cast_fp16 = sigmoid(x = x_591_split_cast_fp16_1)[name = string("x_591_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_591_cast_fp16 = mul(x = x_591_split_cast_fp16_0, y = x_591_split_1_sigmoid_cast_fp16)[name = string("x_591_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_1197_cast_fp16 = select(a = var_45_to_fp16, b = x_591_cast_fp16, cond = var_576)[name = string("input_1197_cast_fp16")];
+            bool new_x_91_interleave_0 = const()[name = string("new_x_91_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_91_cast_fp16 = concat(axis = var_60, interleave = new_x_91_interleave_0, values = (cache_91_cast_fp16, input_1197_cast_fp16))[name = string("new_x_91_cast_fp16")];
+            tensor<int32, [3]> var_5275_begin_0 = const()[name = string("op_5275_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> var_5275_end_0 = const()[name = string("op_5275_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> var_5275_end_mask_0 = const()[name = string("op_5275_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_5275_cast_fp16 = slice_by_index(begin = var_5275_begin_0, end = var_5275_end_0, end_mask = var_5275_end_mask_0, x = new_x_91_cast_fp16)[name = string("op_5275_cast_fp16")];
+            string x_593_pad_type_0 = const()[name = string("x_593_pad_type_0"), val = string("valid")];
+            int32 x_593_groups_0 = const()[name = string("x_593_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_593_strides_0 = const()[name = string("x_593_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_593_pad_0 = const()[name = string("x_593_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_593_dilations_0 = const()[name = string("x_593_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492545984))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492555264))))[name = string("encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_593_cast_fp16 = conv(dilations = x_593_dilations_0, groups = x_593_groups_0, pad = x_593_pad_0, pad_type = x_593_pad_type_0, strides = x_593_strides_0, weight = encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_91_cast_fp16)[name = string("x_593_cast_fp16")];
+            tensor<int32, [3]> input_1199_perm_0 = const()[name = string("input_1199_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_595_axes_0 = const()[name = string("x_595_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492557376)))];
+            tensor<fp16, [1024]> encoder_layers_22_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492559488)))];
+            tensor<fp16, [1, 14, 1024]> input_1199_cast_fp16 = transpose(perm = input_1199_perm_0, x = x_593_cast_fp16)[name = string("transpose_158")];
+            tensor<fp16, [1, 14, 1024]> x_595_cast_fp16 = layer_norm(axes = x_595_axes_0, beta = encoder_layers_22_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_conv_batch_norm_weight_to_fp16, x = input_1199_cast_fp16)[name = string("x_595_cast_fp16")];
+            tensor<int32, [3]> input_1201_perm_0 = const()[name = string("input_1201_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_1201_cast_fp16 = transpose(perm = input_1201_perm_0, x = x_595_cast_fp16)[name = string("transpose_157")];
+            tensor<fp16, [1, 1024, 14]> input_1203_cast_fp16 = silu(x = input_1201_cast_fp16)[name = string("input_1203_cast_fp16")];
+            string x_597_pad_type_0 = const()[name = string("x_597_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_597_strides_0 = const()[name = string("x_597_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_597_pad_0 = const()[name = string("x_597_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_597_dilations_0 = const()[name = string("x_597_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_597_groups_0 = const()[name = string("x_597_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492561600))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493610240))))[name = string("encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_597_cast_fp16 = conv(dilations = x_597_dilations_0, groups = x_597_groups_0, pad = x_597_pad_0, pad_type = x_597_pad_type_0, strides = x_597_strides_0, weight = encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1203_cast_fp16)[name = string("x_597_cast_fp16")];
+            tensor<int32, [3]> input_1205_perm_0 = const()[name = string("input_1205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_1205_cast_fp16 = transpose(perm = input_1205_perm_0, x = x_597_cast_fp16)[name = string("transpose_156")];
+            tensor<fp16, [1, 14, 1024]> input_1207_cast_fp16 = add(x = input_1191_cast_fp16, y = input_1205_cast_fp16)[name = string("input_1207_cast_fp16")];
+            tensor<int32, [1]> input_1209_axes_0 = const()[name = string("input_1209_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493612352)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493614464)))];
+            tensor<fp16, [1, 14, 1024]> input_1209_cast_fp16 = layer_norm(axes = input_1209_axes_0, beta = encoder_layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_feed_forward2_weight_to_fp16, x = input_1207_cast_fp16)[name = string("input_1209_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_22_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493616576)))];
+            tensor<fp16, [4096]> encoder_layers_22_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502005248)))];
+            tensor<fp16, [1, 14, 4096]> linear_206_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear1_weight_to_fp16, x = input_1209_cast_fp16)[name = string("linear_206_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1213_cast_fp16 = silu(x = linear_206_cast_fp16)[name = string("input_1213_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_22_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502013504)))];
+            tensor<fp16, [1024]> encoder_layers_22_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510402176)))];
+            tensor<fp16, [1, 14, 1024]> linear_207_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear2_weight_to_fp16, x = input_1213_cast_fp16)[name = string("linear_207_cast_fp16")];
+            fp16 var_5318_to_fp16 = const()[name = string("op_5318_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_5319_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_5318_to_fp16)[name = string("op_5319_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1219_cast_fp16 = add(x = input_1207_cast_fp16, y = var_5319_cast_fp16)[name = string("input_1219_cast_fp16")];
+            tensor<int32, [1]> input_1221_axes_0 = const()[name = string("input_1221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510404288)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510406400)))];
+            tensor<fp16, [1, 14, 1024]> input_1221_cast_fp16 = layer_norm(axes = input_1221_axes_0, beta = encoder_layers_22_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_22_norm_out_weight_to_fp16, x = input_1219_cast_fp16)[name = string("input_1221_cast_fp16")];
+            tensor<int32, [4]> cache_93_begin_0 = const()[name = string("cache_93_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
+            tensor<int32, [4]> cache_93_end_0 = const()[name = string("cache_93_end_0"), val = tensor<int32, [4]>([24, 1, 42, 1024])];
+            tensor<bool, [4]> cache_93_end_mask_0 = const()[name = string("cache_93_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_93_squeeze_mask_0 = const()[name = string("cache_93_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_93_cast_fp16 = slice_by_index(begin = cache_93_begin_0, end = cache_93_end_0, end_mask = cache_93_end_mask_0, squeeze_mask = cache_93_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_93_cast_fp16")];
+            tensor<int32, [4]> cache_begin_0 = const()[name = string("cache_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
+            tensor<int32, [4]> cache_end_0 = const()[name = string("cache_end_0"), val = tensor<int32, [4]>([24, 1, 1024, 8])];
+            tensor<bool, [4]> cache_end_mask_0 = const()[name = string("cache_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_squeeze_mask_0 = const()[name = string("cache_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_cast_fp16 = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_cast_fp16")];
+            tensor<int32, [1]> input_1223_axes_0 = const()[name = string("input_1223_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510408512)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510410624)))];
+            tensor<fp16, [1, 14, 1024]> input_1223_cast_fp16 = layer_norm(axes = input_1223_axes_0, beta = encoder_layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_feed_forward1_weight_to_fp16, x = input_1221_cast_fp16)[name = string("input_1223_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_23_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510412736)))];
+            tensor<fp16, [4096]> encoder_layers_23_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518801408)))];
+            tensor<fp16, [1, 14, 4096]> linear_208_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear1_weight_to_fp16, x = input_1223_cast_fp16)[name = string("linear_208_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1227_cast_fp16 = silu(x = linear_208_cast_fp16)[name = string("input_1227_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_23_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518809664)))];
+            tensor<fp16, [1024]> encoder_layers_23_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527198336)))];
+            tensor<fp16, [1, 14, 1024]> linear_209_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear2_weight_to_fp16, x = input_1227_cast_fp16)[name = string("linear_209_cast_fp16")];
+            fp16 var_5355_to_fp16 = const()[name = string("op_5355_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_5356_cast_fp16 = mul(x = linear_209_cast_fp16, y = var_5355_to_fp16)[name = string("op_5356_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1233_cast_fp16 = add(x = input_1221_cast_fp16, y = var_5356_cast_fp16)[name = string("input_1233_cast_fp16")];
+            tensor<int32, [1]> key_axes_0 = const()[name = string("key_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527200448)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527202560)))];
+            tensor<fp16, [1, 14, 1024]> key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_23_norm_self_att_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_self_att_weight_to_fp16, x = input_1233_cast_fp16)[name = string("key_cast_fp16")];
+            bool input_1235_interleave_0 = const()[name = string("input_1235_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> input_1235_cast_fp16 = concat(axis = var_69, interleave = input_1235_interleave_0, values = (cache_93_cast_fp16, key_cast_fp16))[name = string("input_1235_cast_fp16")];
+            tensor<int32, [3]> var_5378_begin_0 = const()[name = string("op_5378_begin_0"), val = tensor<int32, [3]>([0, 14, 0])];
+            tensor<int32, [3]> var_5378_end_0 = const()[name = string("op_5378_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_5378_end_mask_0 = const()[name = string("op_5378_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_5378_cast_fp16 = slice_by_index(begin = var_5378_begin_0, end = var_5378_end_0, end_mask = var_5378_end_mask_0, x = cache_93_cast_fp16)[name = string("op_5378_cast_fp16")];
+            bool cache_last_channel_cur_interleave_0 = const()[name = string("cache_last_channel_cur_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> cache_last_channel_cur_cast_fp16 = concat(axis = var_69, interleave = cache_last_channel_cur_interleave_0, values = (var_5378_cast_fp16, key_cast_fp16))[name = string("cache_last_channel_cur_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527204672)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529301888)))];
+            tensor<fp16, [1, 14, 1024]> linear_210_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_q_weight_to_fp16, x = key_cast_fp16)[name = string("linear_210_cast_fp16")];
+            tensor<int32, [4]> var_5389 = const()[name = string("op_5389"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 14, 8, 128]> q_139_cast_fp16 = reshape(shape = var_5389, x = linear_210_cast_fp16)[name = string("q_139_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529304000)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531401216)))];
+            tensor<fp16, [1, 56, 1024]> linear_211_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_k_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_211_cast_fp16")];
+            tensor<int32, [4]> var_5394 = const()[name = string("op_5394"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> k_93_cast_fp16 = reshape(shape = var_5394, x = linear_211_cast_fp16)[name = string("k_93_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531403328)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533500544)))];
+            tensor<fp16, [1, 56, 1024]> linear_212_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_v_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_212_cast_fp16")];
+            tensor<int32, [4]> var_5399 = const()[name = string("op_5399"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> v_cast_fp16 = reshape(shape = var_5399, x = linear_212_cast_fp16)[name = string("v_cast_fp16")];
+            tensor<int32, [4]> value_perm_0 = const()[name = string("value_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_23_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533502656)))];
+            tensor<fp16, [1, 14, 8, 128]> var_5412_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_u_to_fp16)[name = string("op_5412_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_23_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533504768)))];
+            tensor<fp16, [1, 14, 8, 128]> var_5414_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_v_to_fp16)[name = string("op_5414_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_perm_0 = const()[name = string("q_with_bias_v_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_605_transpose_x_0 = const()[name = string("x_605_transpose_x_0"), val = bool(false)];
+            bool x_605_transpose_y_0 = const()[name = string("x_605_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 111]> op_5416_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533506880))), scale = tensor<fp16, [1, 1, 1, 111]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533620608))))[name = string("op_5416_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 14, 128]> q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_5414_cast_fp16)[name = string("transpose_155")];
+            tensor<fp16, [1, 8, 14, 111]> x_605_cast_fp16 = matmul(transpose_x = x_605_transpose_x_0, transpose_y = x_605_transpose_y_0, x = q_with_bias_v_cast_fp16, y = op_5416_to_fp16_quantized)[name = string("x_605_cast_fp16")];
+            tensor<int32, [8]> x_607_pad_0 = const()[name = string("x_607_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_607_mode_0 = const()[name = string("x_607_mode_0"), val = string("constant")];
+            fp16 const_378_to_fp16 = const()[name = string("const_378_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 14, 112]> x_607_cast_fp16 = pad(constant_val = const_378_to_fp16, mode = x_607_mode_0, pad = x_607_pad_0, x = x_605_cast_fp16)[name = string("x_607_cast_fp16")];
+            tensor<int32, [4]> var_5424 = const()[name = string("op_5424"), val = tensor<int32, [4]>([1, 8, -1, 14])];
+            tensor<fp16, [1, 8, 112, 14]> x_609_cast_fp16 = reshape(shape = var_5424, x = x_607_cast_fp16)[name = string("x_609_cast_fp16")];
+            tensor<int32, [4]> var_5428_begin_0 = const()[name = string("op_5428_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5428_end_0 = const()[name = string("op_5428_end_0"), val = tensor<int32, [4]>([1, 8, 112, 14])];
+            tensor<bool, [4]> var_5428_end_mask_0 = const()[name = string("op_5428_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 111, 14]> var_5428_cast_fp16 = slice_by_index(begin = var_5428_begin_0, end = var_5428_end_0, end_mask = var_5428_end_mask_0, x = x_609_cast_fp16)[name = string("op_5428_cast_fp16")];
+            tensor<int32, [4]> var_5429 = const()[name = string("op_5429"), val = tensor<int32, [4]>([1, 8, 14, 111])];
+            tensor<fp16, [1, 8, 14, 111]> matrix_bd_93_cast_fp16 = reshape(shape = var_5429, x = var_5428_cast_fp16)[name = string("matrix_bd_93_cast_fp16")];
+            bool matrix_ac_transpose_x_0 = const()[name = string("matrix_ac_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_transpose_y_0 = const()[name = string("matrix_ac_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_142_perm_0 = const()[name = string("transpose_142_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_143_perm_0 = const()[name = string("transpose_143_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 56]> transpose_143 = transpose(perm = transpose_143_perm_0, x = k_93_cast_fp16)[name = string("transpose_153")];
+            tensor<fp16, [1, 8, 14, 128]> transpose_142 = transpose(perm = transpose_142_perm_0, x = var_5412_cast_fp16)[name = string("transpose_154")];
+            tensor<fp16, [1, 8, 14, 56]> matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_142, y = transpose_143)[name = string("matrix_ac_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_begin_0 = const()[name = string("matrix_bd_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_end_0 = const()[name = string("matrix_bd_end_0"), val = tensor<int32, [4]>([1, 8, 14, 56])];
+            tensor<bool, [4]> matrix_bd_end_mask_0 = const()[name = string("matrix_bd_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 14, 56]> matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_93_cast_fp16)[name = string("matrix_bd_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_5438_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = string("op_5438_cast_fp16")];
+            fp16 _inversed_scores_93_y_0_to_fp16 = const()[name = string("_inversed_scores_93_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 14, 56]> _inversed_scores_93_cast_fp16 = mul(x = var_5438_cast_fp16, y = _inversed_scores_93_y_0_to_fp16)[name = string("_inversed_scores_93_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> scores_cast_fp16 = select(a = var_46_to_fp16, b = _inversed_scores_93_cast_fp16, cond = mask_11)[name = string("scores_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> var_5444_cast_fp16 = softmax(axis = var_60, x = scores_cast_fp16)[name = string("op_5444_cast_fp16")];
+            tensor<fp16, [1, 8, 14, 56]> input_1237_cast_fp16 = select(a = var_45_to_fp16, b = var_5444_cast_fp16, cond = mask_11)[name = string("input_1237_cast_fp16")];
+            bool x_611_transpose_x_0 = const()[name = string("x_611_transpose_x_0"), val = bool(false)];
+            bool x_611_transpose_y_0 = const()[name = string("x_611_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 56, 128]> value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = string("transpose_152")];
+            tensor<fp16, [1, 8, 14, 128]> x_611_cast_fp16 = matmul(transpose_x = x_611_transpose_x_0, transpose_y = x_611_transpose_y_0, x = input_1237_cast_fp16, y = value_cast_fp16)[name = string("x_611_cast_fp16")];
+            tensor<int32, [4]> var_5448_perm_0 = const()[name = string("op_5448_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5449 = const()[name = string("op_5449"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 14, 8, 128]> var_5448_cast_fp16 = transpose(perm = var_5448_perm_0, x = x_611_cast_fp16)[name = string("transpose_151")];
+            tensor<fp16, [1, 14, 1024]> input_1239_cast_fp16 = reshape(shape = var_5449, x = var_5448_cast_fp16)[name = string("input_1239_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533620928)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535718144)))];
+            tensor<fp16, [1, 14, 1024]> linear_214_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_out_weight_to_fp16, x = input_1239_cast_fp16)[name = string("linear_214_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1243_cast_fp16 = add(x = input_1233_cast_fp16, y = linear_214_cast_fp16)[name = string("input_1243_cast_fp16")];
+            tensor<int32, [1]> x_615_axes_0 = const()[name = string("x_615_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535720256)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535722368)))];
+            tensor<fp16, [1, 14, 1024]> x_615_cast_fp16 = layer_norm(axes = x_615_axes_0, beta = encoder_layers_23_norm_conv_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_conv_weight_to_fp16, x = input_1243_cast_fp16)[name = string("x_615_cast_fp16")];
+            tensor<int32, [3]> input_1245_perm_0 = const()[name = string("input_1245_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1247_pad_type_0 = const()[name = string("input_1247_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1247_strides_0 = const()[name = string("input_1247_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1247_pad_0 = const()[name = string("input_1247_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1247_dilations_0 = const()[name = string("input_1247_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1247_groups_0 = const()[name = string("input_1247_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535724480))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537821696))))[name = string("encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> input_1245_cast_fp16 = transpose(perm = input_1245_perm_0, x = x_615_cast_fp16)[name = string("transpose_150")];
+            tensor<fp16, [1, 2048, 14]> input_1247_cast_fp16 = conv(dilations = input_1247_dilations_0, groups = input_1247_groups_0, pad = input_1247_pad_0, pad_type = input_1247_pad_type_0, strides = input_1247_strides_0, weight = encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1245_cast_fp16)[name = string("input_1247_cast_fp16")];
+            int32 x_617_split_num_splits_0 = const()[name = string("x_617_split_num_splits_0"), val = int32(2)];
+            int32 x_617_split_axis_0 = const()[name = string("x_617_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 14]> x_617_split_cast_fp16_0, tensor<fp16, [1, 1024, 14]> x_617_split_cast_fp16_1 = split(axis = x_617_split_axis_0, num_splits = x_617_split_num_splits_0, x = input_1247_cast_fp16)[name = string("x_617_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_617_split_1_sigmoid_cast_fp16 = sigmoid(x = x_617_split_cast_fp16_1)[name = string("x_617_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> x_617_cast_fp16 = mul(x = x_617_split_cast_fp16_0, y = x_617_split_1_sigmoid_cast_fp16)[name = string("x_617_cast_fp16")];
+            tensor<fp16, [1, 1024, 14]> input_1249_cast_fp16 = select(a = var_45_to_fp16, b = x_617_cast_fp16, cond = var_576)[name = string("input_1249_cast_fp16")];
+            bool new_x_interleave_0 = const()[name = string("new_x_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 22]> new_x_cast_fp16 = concat(axis = var_60, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_1249_cast_fp16))[name = string("new_x_cast_fp16")];
+            tensor<int32, [3]> cache_last_time_cur_begin_0 = const()[name = string("cache_last_time_cur_begin_0"), val = tensor<int32, [3]>([0, 0, 14])];
+            tensor<int32, [3]> cache_last_time_cur_end_0 = const()[name = string("cache_last_time_cur_end_0"), val = tensor<int32, [3]>([1, 1024, 22])];
+            tensor<bool, [3]> cache_last_time_cur_end_mask_0 = const()[name = string("cache_last_time_cur_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> cache_last_time_cur_cast_fp16 = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = new_x_cast_fp16)[name = string("cache_last_time_cur_cast_fp16")];
+            string x_619_pad_type_0 = const()[name = string("x_619_pad_type_0"), val = string("valid")];
+            int32 x_619_groups_0 = const()[name = string("x_619_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_619_strides_0 = const()[name = string("x_619_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_619_pad_0 = const()[name = string("x_619_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_619_dilations_0 = const()[name = string("x_619_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537825856))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537835136))))[name = string("encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_619_cast_fp16 = conv(dilations = x_619_dilations_0, groups = x_619_groups_0, pad = x_619_pad_0, pad_type = x_619_pad_type_0, strides = x_619_strides_0, weight = encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_cast_fp16)[name = string("x_619_cast_fp16")];
+            tensor<int32, [3]> input_1251_perm_0 = const()[name = string("input_1251_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_621_axes_0 = const()[name = string("x_621_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537837248)))];
+            tensor<fp16, [1024]> encoder_layers_23_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537839360)))];
+            tensor<fp16, [1, 14, 1024]> input_1251_cast_fp16 = transpose(perm = input_1251_perm_0, x = x_619_cast_fp16)[name = string("transpose_149")];
+            tensor<fp16, [1, 14, 1024]> x_621_cast_fp16 = layer_norm(axes = x_621_axes_0, beta = encoder_layers_23_conv_batch_norm_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_conv_batch_norm_weight_to_fp16, x = input_1251_cast_fp16)[name = string("x_621_cast_fp16")];
+            tensor<int32, [3]> input_1253_perm_0 = const()[name = string("input_1253_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 14]> input_1253_cast_fp16 = transpose(perm = input_1253_perm_0, x = x_621_cast_fp16)[name = string("transpose_148")];
+            tensor<fp16, [1, 1024, 14]> input_1255_cast_fp16 = silu(x = input_1253_cast_fp16)[name = string("input_1255_cast_fp16")];
+            string x_623_pad_type_0 = const()[name = string("x_623_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_623_strides_0 = const()[name = string("x_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_623_pad_0 = const()[name = string("x_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_623_dilations_0 = const()[name = string("x_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_623_groups_0 = const()[name = string("x_623_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537841472))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538890112))))[name = string("encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 14]> x_623_cast_fp16 = conv(dilations = x_623_dilations_0, groups = x_623_groups_0, pad = x_623_pad_0, pad_type = x_623_pad_type_0, strides = x_623_strides_0, weight = encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1255_cast_fp16)[name = string("x_623_cast_fp16")];
+            tensor<int32, [3]> input_1257_perm_0 = const()[name = string("input_1257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 14, 1024]> input_1257_cast_fp16 = transpose(perm = input_1257_perm_0, x = x_623_cast_fp16)[name = string("transpose_147")];
+            tensor<fp16, [1, 14, 1024]> input_1259_cast_fp16 = add(x = input_1243_cast_fp16, y = input_1257_cast_fp16)[name = string("input_1259_cast_fp16")];
+            tensor<int32, [1]> input_1261_axes_0 = const()[name = string("input_1261_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538892224)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538894336)))];
+            tensor<fp16, [1, 14, 1024]> input_1261_cast_fp16 = layer_norm(axes = input_1261_axes_0, beta = encoder_layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_feed_forward2_weight_to_fp16, x = input_1259_cast_fp16)[name = string("input_1261_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_23_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538896448)))];
+            tensor<fp16, [4096]> encoder_layers_23_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547285120)))];
+            tensor<fp16, [1, 14, 4096]> linear_215_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear1_weight_to_fp16, x = input_1261_cast_fp16)[name = string("linear_215_cast_fp16")];
+            tensor<fp16, [1, 14, 4096]> input_1265_cast_fp16 = silu(x = linear_215_cast_fp16)[name = string("input_1265_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_23_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547293376)))];
+            tensor<fp16, [1024]> encoder_layers_23_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555682048)))];
+            tensor<fp16, [1, 14, 1024]> linear_216_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear2_weight_to_fp16, x = input_1265_cast_fp16)[name = string("linear_216_cast_fp16")];
+            fp16 var_5531_to_fp16 = const()[name = string("op_5531_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 14, 1024]> var_5532_cast_fp16 = mul(x = linear_216_cast_fp16, y = var_5531_to_fp16)[name = string("op_5532_cast_fp16")];
+            tensor<fp16, [1, 14, 1024]> input_1271_cast_fp16 = add(x = input_1259_cast_fp16, y = var_5532_cast_fp16)[name = string("input_1271_cast_fp16")];
+            tensor<int32, [1]> audio_signal_axes_0 = const()[name = string("audio_signal_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555684160)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555686272)))];
+            tensor<fp16, [1, 14, 1024]> audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_23_norm_out_bias_to_fp16, epsilon = var_43_to_fp16, gamma = encoder_layers_23_norm_out_weight_to_fp16, x = input_1271_cast_fp16)[name = string("audio_signal_cast_fp16")];
+            int32 obj_5_axis_0 = const()[name = string("obj_5_axis_0"), val = int32(0)];
+            tensor<fp16, [24, 1, 42, 1024]> obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_485_cast_fp16, var_698_cast_fp16, var_911_cast_fp16, var_1124_cast_fp16, var_1337_cast_fp16, var_1550_cast_fp16, var_1763_cast_fp16, var_1976_cast_fp16, var_2189_cast_fp16, var_2402_cast_fp16, var_2615_cast_fp16, var_2828_cast_fp16, var_3041_cast_fp16, var_3254_cast_fp16, var_3467_cast_fp16, var_3680_cast_fp16, var_3893_cast_fp16, var_4106_cast_fp16, var_4319_cast_fp16, var_4532_cast_fp16, var_4745_cast_fp16, var_4958_cast_fp16, var_5171_cast_fp16, cache_last_channel_cur_cast_fp16))[name = string("obj_5_cast_fp16")];
+            int32 obj_7_axis_0 = const()[name = string("obj_7_axis_0"), val = int32(0)];
+            tensor<fp16, [24, 1, 1024, 8]> obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_589_cast_fp16, var_802_cast_fp16, var_1015_cast_fp16, var_1228_cast_fp16, var_1441_cast_fp16, var_1654_cast_fp16, var_1867_cast_fp16, var_2080_cast_fp16, var_2293_cast_fp16, var_2506_cast_fp16, var_2719_cast_fp16, var_2932_cast_fp16, var_3145_cast_fp16, var_3358_cast_fp16, var_3571_cast_fp16, var_3784_cast_fp16, var_3997_cast_fp16, var_4210_cast_fp16, var_4423_cast_fp16, var_4636_cast_fp16, var_4849_cast_fp16, var_5062_cast_fp16, var_5275_cast_fp16, cache_last_time_cur_cast_fp16))[name = string("obj_7_cast_fp16")];
+            tensor<int32, [1]> var_5548 = add(x = cache_len, y = max_audio_length_1)[name = string("op_5548")];
+            string var_5548_promoted_to_fp16_dtype_0 = const()[name = string("op_5548_promoted_to_fp16_dtype_0"), val = string("fp16")];
+            fp16 const_384_to_fp16 = const()[name = string("const_384_to_fp16"), val = fp16(-inf)];
+            fp16 var_50_promoted_to_fp16 = const()[name = string("op_50_promoted_to_fp16"), val = fp16(0x1.5p+5)];
+            tensor<fp16, [1]> var_5548_to_fp16 = cast(dtype = var_5548_promoted_to_fp16_dtype_0, x = var_5548)[name = string("cast_10")];
+            tensor<fp16, [1]> clip_1_cast_fp16 = clip(alpha = const_384_to_fp16, beta = var_50_promoted_to_fp16, x = var_5548_to_fp16)[name = string("clip_1_cast_fp16")];
+            int32 one_hot_1_batch_dims_0 = const()[name = string("one_hot_1_batch_dims_0"), val = int32(0)];
+            bool one_hot_1_validate_indices_0 = const()[name = string("one_hot_1_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [128, 128]> to_onehot_identity_table_to_fp16 = const()[name = string("to_onehot_identity_table_to_fp16"), val = tensor<fp16, [128, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555688384)))];
+            string prompt_id_to_int16_dtype_0 = const()[name = string("prompt_id_to_int16_dtype_0"), val = string("int16")];
+            string cast_230_dtype_0 = const()[name = string("cast_230_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
+            tensor<int16, [1]> prompt_id_to_int16 = cast(dtype = prompt_id_to_int16_dtype_0, x = prompt_id)[name = string("cast_9")];
+            tensor<int32, [1]> cast_230 = cast(dtype = cast_230_dtype_0, x = prompt_id_to_int16)[name = string("cast_8")];
+            tensor<bool, [1]> greater_equal_0 = greater_equal(x = cast_230, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
+            int32 slice_by_index_2 = const()[name = string("slice_by_index_2"), val = int32(128)];
+            tensor<int32, [1]> add_0 = add(x = cast_230, y = slice_by_index_2)[name = string("add_0")];
+            tensor<int32, [1]> select_0 = select(a = cast_230, b = add_0, cond = greater_equal_0)[name = string("select_0")];
+            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
+            string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
+            tensor<int16, [1]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_7")];
+            tensor<int32, [1]> cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_6")];
+            tensor<bool, [1]> greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
+            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(128)];
+            tensor<int32, [1]> add_0_1 = add(x = cast_0, y = slice_by_index_0)[name = string("add_0_1")];
+            tensor<int32, [1]> select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")];
+            int32 greater_equal_0_y_0_2 = const()[name = string("greater_equal_0_y_0_2"), val = int32(0)];
+            tensor<bool, [1]> greater_equal_0_2 = greater_equal(x = select_0_1, y = greater_equal_0_y_0_2)[name = string("greater_equal_0_2")];
+            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(128)];
+            tensor<int32, [1]> add_0_2 = add(x = select_0_1, y = slice_by_index_0_1)[name = string("add_0_2")];
+            tensor<int32, [1]> select_0_2 = select(a = select_0_1, b = add_0_2, cond = greater_equal_0_2)[name = string("select_0_2")];
+            int32 one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(0)];
+            tensor<fp16, [1, 128]> one_hot_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = one_hot_1_batch_dims_0, indices = select_0_2, validate_indices = one_hot_1_validate_indices_0, x = to_onehot_identity_table_to_fp16)[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16")];
+            tensor<int32, [1]> var_5594_axes_0 = const()[name = string("op_5594_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 128]> var_5594_cast_fp16 = expand_dims(axes = var_5594_axes_0, x = one_hot_1_cast_fp16_cast_uint16_cast_uint16)[name = string("op_5594_cast_fp16")];
+            tensor<int32, [3]> one_hot_reps_0 = const()[name = string("one_hot_reps_0"), val = tensor<int32, [3]>([1, 14, 1])];
+            tensor<fp16, [1, 14, 128]> one_hot_cast_fp16 = tile(reps = one_hot_reps_0, x = var_5594_cast_fp16)[name = string("one_hot_cast_fp16")];
+            int32 var_5603 = const()[name = string("op_5603"), val = int32(-1)];
+            bool input_1273_interleave_0 = const()[name = string("input_1273_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 14, 1152]> input_1273_cast_fp16 = concat(axis = var_5603, interleave = input_1273_interleave_0, values = (audio_signal_cast_fp16, one_hot_cast_fp16))[name = string("input_1273_cast_fp16")];
+            tensor<fp16, [2048, 1152]> prompt_kernel_0_weight_to_fp16 = const()[name = string("prompt_kernel_0_weight_to_fp16"), val = tensor<fp16, [2048, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555721216)))];
+            tensor<fp16, [2048]> prompt_kernel_0_bias_to_fp16 = const()[name = string("prompt_kernel_0_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560439872)))];
+            tensor<fp16, [1, 14, 2048]> linear_217_cast_fp16 = linear(bias = prompt_kernel_0_bias_to_fp16, weight = prompt_kernel_0_weight_to_fp16, x = input_1273_cast_fp16)[name = string("linear_217_cast_fp16")];
+            tensor<fp16, [1, 14, 2048]> input_1277_cast_fp16 = relu(x = linear_217_cast_fp16)[name = string("input_1277_cast_fp16")];
+            tensor<fp16, [1024, 2048]> prompt_kernel_2_weight_to_fp16 = const()[name = string("prompt_kernel_2_weight_to_fp16"), val = tensor<fp16, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560444032)))];
+            tensor<fp16, [1024]> prompt_kernel_2_bias_to_fp16 = const()[name = string("prompt_kernel_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564638400)))];
+            tensor<fp16, [1, 14, 1024]> linear_218_cast_fp16 = linear(bias = prompt_kernel_2_bias_to_fp16, weight = prompt_kernel_2_weight_to_fp16, x = input_1277_cast_fp16)[name = string("linear_218_cast_fp16")];
+            string conditioned_cast_fp16_to_fp32_dtype_0 = const()[name = string("conditioned_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [3]> transpose_72_perm_0_1 = const()[name = string("transpose_72_perm_0_1"), val = tensor<int32, [3]>([0, 2, 1])];
+            string var_5621_dtype_0 = const()[name = string("op_5621_dtype_0"), val = string("int32")];
+            tensor<int32, [4]> var_5624_perm_0 = const()[name = string("op_5624_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string var_5624_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5624_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [4]> var_5627_perm_0 = const()[name = string("op_5627_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string var_5627_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5627_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            string var_5632_dtype_0 = const()[name = string("op_5632_dtype_0"), val = string("int32")];
+            tensor<fp16, [640, 1024]> joint_enc_weight_to_fp16 = const()[name = string("joint_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564640512)))];
+            tensor<fp16, [640]> joint_enc_bias_to_fp16 = const()[name = string("joint_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565951296)))];
+            tensor<fp16, [1, 14, 640]> linear_219_cast_fp16 = linear(bias = joint_enc_bias_to_fp16, weight = joint_enc_weight_to_fp16, x = linear_218_cast_fp16)[name = string("linear_219_cast_fp16")];
+            string linear_219_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_219_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 14, 640]> encoder_proj = cast(dtype = linear_219_cast_fp16_to_fp32_dtype_0, x = linear_219_cast_fp16)[name = string("cast_0")];
+            tensor<int32, [1]> cache_len_out = cast(dtype = var_5632_dtype_0, x = clip_1_cast_fp16)[name = string("cast_1")];
+            tensor<fp16, [1, 24, 1024, 8]> var_5627_cast_fp16 = transpose(perm = var_5627_perm_0, x = obj_7_cast_fp16)[name = string("transpose_144")];
+            tensor<fp32, [1, 24, 1024, 8]> cache_time_out = cast(dtype = var_5627_cast_fp16_to_fp32_dtype_0, x = var_5627_cast_fp16)[name = string("cast_2")];
+            tensor<fp16, [1, 24, 42, 1024]> var_5624_cast_fp16 = transpose(perm = var_5624_perm_0, x = obj_5_cast_fp16)[name = string("transpose_145")];
+            tensor<fp32, [1, 24, 42, 1024]> cache_channel_out = cast(dtype = var_5624_cast_fp16_to_fp32_dtype_0, x = var_5624_cast_fp16)[name = string("cast_3")];
+            tensor<int32, [1]> encoded_length = cast(dtype = var_5621_dtype_0, x = clip_0_cast_fp16)[name = string("cast_4")];
+            tensor<fp16, [1, 1024, 14]> transpose_72_1 = transpose(perm = transpose_72_perm_0_1, x = linear_218_cast_fp16)[name = string("transpose_146")];
+            tensor<fp32, [1, 1024, 14]> encoded = cast(dtype = conditioned_cast_fp16_to_fp32_dtype_0, x = transpose_72_1)[name = string("cast_5")];
+        } -> (encoded, encoded_length, cache_channel_out, cache_time_out, cache_len_out, encoder_proj);
+}
\ No newline at end of file
diff --git a/fr/1120ms/encoder.mlmodelc/weights/weight.bin b/fr/1120ms/encoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..db719d65717ba13a2abcc1fd6915682c3eeef841
--- /dev/null
+++ b/fr/1120ms/encoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:011a5342cc1e2901633a8f0468be561f50dfc289ebe51851cddb4831f9a6a23f
+size 565952640
diff --git a/fr/1120ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/1120ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..d7e63d6d9ce257a9bcf94ee5f3694e8fe97c90ab
--- /dev/null
+++ b/fr/1120ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14655ea0dbccc666356247f05afc87bbea77ff469344bfb084526e00c0ce6d15
+size 804512
diff --git a/fr/1120ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/1120ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..db719d65717ba13a2abcc1fd6915682c3eeef841
--- /dev/null
+++ b/fr/1120ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:011a5342cc1e2901633a8f0468be561f50dfc289ebe51851cddb4831f9a6a23f
+size 565952640
diff --git a/fr/1120ms/encoder.mlpackage/Manifest.json b/fr/1120ms/encoder.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..bba3892a42b16ed56b063a7a2b710b5148d36277
--- /dev/null
+++ b/fr/1120ms/encoder.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "6BA56D30-2135-407A-9AD7-D9996C9B4325": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "D3843811-369C-4145-9609-E516DF19D3E5": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "6BA56D30-2135-407A-9AD7-D9996C9B4325"
+}
diff --git a/fr/1120ms/joint.mlmodelc/analytics/coremldata.bin b/fr/1120ms/joint.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..38aade026cc9376471863f9f09be292c9fe8578b
--- /dev/null
+++ b/fr/1120ms/joint.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad1f6f5571ee9e7ad8a533b4679ed7416879f4d8e5fc10f40835c1d014b822a4
+size 243
diff --git a/fr/1120ms/joint.mlmodelc/coremldata.bin b/fr/1120ms/joint.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..45df627e7953068f36ea8f3e5467664cf318a44e
--- /dev/null
+++ b/fr/1120ms/joint.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e630dfefcd0f652f2815f76cfbacb89b51f2c94618fbc517e41779426099a1f
+size 341
diff --git a/fr/1120ms/joint.mlmodelc/model.mil b/fr/1120ms/joint.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..041dd87f66df1673f6226d5c642c93a64d4e2730
--- /dev/null
+++ b/fr/1120ms/joint.mlmodelc/model.mil
@@ -0,0 +1,31 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp32, [1, 640, 1]> decoder, tensor<fp32, [1, 1024, 1]> encoder) {
+            tensor<int32, [3]> input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [640, 1024]> module_enc_weight_to_fp16 = const()[name = string("module_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [640]> module_enc_bias_to_fp16 = const()[name = string("module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1310848)))];
+            tensor<fp16, [1, 1024, 1]> encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_2")];
+            tensor<fp16, [1, 1, 1024]> input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_to_fp16)[name = string("transpose_1")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = module_enc_bias_to_fp16, weight = module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> module_pred_weight_to_fp16 = const()[name = string("module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1312192)))];
+            tensor<fp16, [640]> module_pred_bias_to_fp16 = const()[name = string("module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2131456)))];
+            tensor<fp16, [1, 640, 1]> decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_1")];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_to_fp16)[name = string("transpose_0")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = module_pred_bias_to_fp16, weight = module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_23_axes_0 = const()[name = string("op_23_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = string("op_23_cast_fp16")];
+            tensor<int32, [1]> var_25_axes_0 = const()[name = string("op_25_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_25_cast_fp16 = expand_dims(axes = var_25_axes_0, x = linear_1_cast_fp16)[name = string("op_25_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_25_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("input_7_cast_fp16")];
+            tensor<fp16, [849, 640]> module_joint_net_2_weight_to_fp16 = const()[name = string("module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2132800)))];
+            tensor<fp16, [849]> module_joint_net_2_bias_to_fp16 = const()[name = string("module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [849]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3219584)))];
+            tensor<fp16, [1, 1, 1, 849]> linear_2_cast_fp16 = linear(bias = module_joint_net_2_bias_to_fp16, weight = module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_2_cast_fp16")];
+            string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 1, 1, 849]> logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_0")];
+        } -> (logits);
+}
\ No newline at end of file
diff --git a/fr/1120ms/joint.mlmodelc/weights/weight.bin b/fr/1120ms/joint.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..109a5bf4eb27d11fcd040ea350a8e6d00d18932a
--- /dev/null
+++ b/fr/1120ms/joint.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9fa5fb0d14d87484d9ffd1f4d30996ea456c14ea3e5ab9d910486d8e86d3a540
+size 3221346
diff --git a/fr/1120ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/1120ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..0c650fe487d18d0d937a1fc295a3eee53b4aacbe
--- /dev/null
+++ b/fr/1120ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4bac39a8a21b3e7e8cdeba1d03b92e8f743bf87cb2de3b1d85c505a580a0b218
+size 4486
diff --git a/fr/1120ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/1120ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..109a5bf4eb27d11fcd040ea350a8e6d00d18932a
--- /dev/null
+++ b/fr/1120ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9fa5fb0d14d87484d9ffd1f4d30996ea456c14ea3e5ab9d910486d8e86d3a540
+size 3221346
diff --git a/fr/1120ms/joint.mlpackage/Manifest.json b/fr/1120ms/joint.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7e49a7390ba8e327aff397061f0fd4476b7c70a
--- /dev/null
+++ b/fr/1120ms/joint.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "0DC18C6B-B3DE-4D3A-80EE-28CAA06F6708": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "88105E4A-1C49-40F4-A835-C9B725CEEC59": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "88105E4A-1C49-40F4-A835-C9B725CEEC59"
+}
diff --git a/fr/1120ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin b/fr/1120ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..a856ad184cad37664eb30bb699b3a88e6401000d
--- /dev/null
+++ b/fr/1120ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28b5c13f8ce2d8154edb31968c174d322ee4ae76db42f41ecf7ac5e0fd87f57d
+size 243
diff --git a/fr/1120ms/joint_noencproj_batched.mlmodelc/coremldata.bin b/fr/1120ms/joint_noencproj_batched.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..b68dc2ebdbe0b8957cd45ed4b265fda6b6195204
--- /dev/null
+++ b/fr/1120ms/joint_noencproj_batched.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5dc74778713432d34163ae80901d3294eef3341dd020a29dc05d67eba8f6f0a
+size 406
diff --git a/fr/1120ms/joint_noencproj_batched.mlmodelc/model.mil b/fr/1120ms/joint_noencproj_batched.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..b66479ae4eaa92652a98a34a18e993259a14e6c0
--- /dev/null
+++ b/fr/1120ms/joint_noencproj_batched.mlmodelc/model.mil
@@ -0,0 +1,26 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios18>(tensor<fp32, [1, 640, 1]> decoder, tensor<fp32, [1, 4, 640]> encoder_proj) {
+            tensor<int32, [3]> input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819328)))];
+            tensor<fp16, [1, 640, 1]> decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_2")];
+            tensor<fp16, [1, 1, 640]> input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = decoder_to_fp16)[name = string("transpose_0")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [1]> var_15_axes_0 = const()[name = string("op_15_axes_0"), val = tensor<int32, [1]>([2])];
+            string encoder_proj_to_fp16_dtype_0 = const()[name = string("encoder_proj_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 4, 640]> encoder_proj_to_fp16 = cast(dtype = encoder_proj_to_fp16_dtype_0, x = encoder_proj)[name = string("cast_1")];
+            tensor<fp16, [1, 4, 1, 640]> var_15_cast_fp16 = expand_dims(axes = var_15_axes_0, x = encoder_proj_to_fp16)[name = string("op_15_cast_fp16")];
+            tensor<int32, [1]> var_17_axes_0 = const()[name = string("op_17_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_17_cast_fp16 = expand_dims(axes = var_17_axes_0, x = linear_0_cast_fp16)[name = string("op_17_cast_fp16")];
+            tensor<fp16, [1, 4, 1, 640]> input_3_cast_fp16 = add(x = var_15_cast_fp16, y = var_17_cast_fp16)[name = string("input_3_cast_fp16")];
+            tensor<fp16, [1, 4, 1, 640]> input_5_cast_fp16 = relu(x = input_3_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<fp16, [849, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(820672)))];
+            tensor<fp16, [849]> joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [849]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1907456)))];
+            tensor<fp16, [1, 4, 1, 849]> linear_1_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_5_cast_fp16)[name = string("linear_1_cast_fp16")];
+            string linear_1_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_1_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 4, 1, 849]> logits = cast(dtype = linear_1_cast_fp16_to_fp32_dtype_0, x = linear_1_cast_fp16)[name = string("cast_0")];
+        } -> (logits);
+}
\ No newline at end of file
diff --git a/fr/1120ms/joint_noencproj_batched.mlmodelc/weights/weight.bin b/fr/1120ms/joint_noencproj_batched.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..4e66049a9a8bde2b92859df3192b671ddf9ab81d
--- /dev/null
+++ b/fr/1120ms/joint_noencproj_batched.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9a9dd24e6b63b97540b0db0e990418bc6fddc9c2bf85b82eb226e6df1ea02a4
+size 1909218
diff --git a/fr/1120ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/1120ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..4f82db74749dc3350042c3f6c1453c4d4e620d50
--- /dev/null
+++ b/fr/1120ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d13f08c4961187e47001773d7d4d5d5937ba7ff68475eb7c1bac41d7014c64a1
+size 3839
diff --git a/fr/1120ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/1120ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..4e66049a9a8bde2b92859df3192b671ddf9ab81d
--- /dev/null
+++ b/fr/1120ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9a9dd24e6b63b97540b0db0e990418bc6fddc9c2bf85b82eb226e6df1ea02a4
+size 1909218
diff --git a/fr/1120ms/joint_noencproj_batched.mlpackage/Manifest.json b/fr/1120ms/joint_noencproj_batched.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..5f2e3080d9f6dc5617b9630128af453e31562411
--- /dev/null
+++ b/fr/1120ms/joint_noencproj_batched.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "EFDDA9F1-A088-4CE2-8A64-5928EB8F429C": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "F28FF7B3-AC0A-4ACA-93F8-DF201DFFE75B": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "F28FF7B3-AC0A-4ACA-93F8-DF201DFFE75B"
+}
diff --git a/fr/1120ms/metadata.json b/fr/1120ms/metadata.json
new file mode 100644
index 0000000000000000000000000000000000000000..23b5b5330d4faac993ec1ca8f4fc3cb8ba7f34e2
--- /dev/null
+++ b/fr/1120ms/metadata.json
@@ -0,0 +1,198 @@
+{
+  "model": "nvidia/nemotron-asr-streaming-multilingual-0.6b",
+  "model_class": "nemo.collections.asr.models.rnnt_bpe_models_prompt.EncDecRNNTBPEModelWithPrompt",
+  "sample_rate": 16000,
+  "mel_features": 128,
+  "chunk_mel_frames": 112,
+  "pre_encode_cache": 9,
+  "total_mel_frames": 121,
+  "att_context_size": [
+    42,
+    13
+  ],
+  "vocab_size": 848,
+  "blank_idx": 848,
+  "vocab_pruned": true,
+  "vocab_pruned_original_size": 13087,
+  "cache_channel_shape": [
+    1,
+    24,
+    42,
+    1024
+  ],
+  "cache_time_shape": [
+    1,
+    24,
+    1024,
+    8
+  ],
+  "decoder_hidden": 640,
+  "decoder_layers": 2,
+  "encoder_dim": 1024,
+  "num_prompts": 128,
+  "prompt_dictionary": {
+    "af-ZA": 54,
+    "am-ET": 49,
+    "ar": 7,
+    "ar-AR": 7,
+    "auto": 101,
+    "ay-BO": 81,
+    "az-AZ": 66,
+    "bg": 30,
+    "bg-BG": 30,
+    "bn-IN": 36,
+    "cs": 22,
+    "cs-CZ": 22,
+    "da": 25,
+    "da-DK": 25,
+    "de": 9,
+    "de-DE": 9,
+    "el": 21,
+    "el-GR": 21,
+    "en": 0,
+    "en-GB": 1,
+    "en-US": 0,
+    "enGB": 1,
+    "es": 3,
+    "es-ES": 2,
+    "es-US": 3,
+    "esES": 2,
+    "et": 60,
+    "et-EE": 60,
+    "fa-IR": 38,
+    "fi": 26,
+    "fi-FI": 26,
+    "fr": 8,
+    "fr-CA": 100,
+    "fr-FR": 8,
+    "gn-PY": 82,
+    "gu-IN": 42,
+    "ha-NG": 50,
+    "haw-US": 97,
+    "he-IL": 64,
+    "hi": 6,
+    "hi-HI": 6,
+    "hi-IN": 6,
+    "hr": 29,
+    "hr-HR": 29,
+    "hu": 23,
+    "hu-HU": 23,
+    "hy-AM": 68,
+    "id-ID": 34,
+    "ig-NG": 53,
+    "it": 15,
+    "it-IT": 15,
+    "ja-JA": 10,
+    "ja-JP": 10,
+    "ka-GE": 67,
+    "km-KH": 47,
+    "kn-IN": 43,
+    "ko": 14,
+    "ko-KO": 14,
+    "ko-KR": 14,
+    "ku-TR": 65,
+    "ky-KG": 71,
+    "ln-CD": 58,
+    "lt": 31,
+    "lt-LT": 31,
+    "lv": 61,
+    "lv-LV": 61,
+    "mi-NZ": 96,
+    "ml-IN": 44,
+    "mr-IN": 41,
+    "ms-MY": 35,
+    "mt-MT": 102,
+    "nah-MX": 83,
+    "nb": 103,
+    "nb-NO": 103,
+    "ne-NP": 46,
+    "nl": 16,
+    "nl-NL": 16,
+    "nn": 104,
+    "nn-NO": 104,
+    "no": 27,
+    "no-NO": 27,
+    "ny-MW": 57,
+    "or-KE": 59,
+    "pl": 17,
+    "pl-PL": 17,
+    "pt": 13,
+    "pt-BR": 12,
+    "pt-PT": 13,
+    "qu-PE": 80,
+    "ro": 20,
+    "ro-RO": 20,
+    "ru": 11,
+    "ru-RU": 11,
+    "rw-RW": 55,
+    "si-LK": 45,
+    "sk": 28,
+    "sk-SK": 28,
+    "sl": 62,
+    "sl-SI": 62,
+    "sm-WS": 98,
+    "so-SO": 56,
+    "sv": 24,
+    "sv-SE": 24,
+    "sw-KE": 48,
+    "ta-IN": 39,
+    "te-IN": 40,
+    "tg-TJ": 70,
+    "th-TH": 32,
+    "to-TO": 99,
+    "tr": 18,
+    "tr-TR": 18,
+    "uk": 19,
+    "uk-UA": 19,
+    "ur-PK": 37,
+    "uz-UZ": 69,
+    "vi-VN": 33,
+    "yo-NG": 52,
+    "zh-CN": 4,
+    "zh-TW": 5,
+    "zh-ZH": 4,
+    "zu-ZA": 51
+  },
+  "default_prompt_id": 101,
+  "lang_tag_token_ids": [
+    1,
+    72,
+    122,
+    170,
+    230,
+    231,
+    263,
+    282,
+    390,
+    415,
+    430,
+    483,
+    501,
+    519,
+    542,
+    565,
+    594,
+    619,
+    620,
+    629,
+    640,
+    659,
+    660,
+    677,
+    704,
+    717,
+    735,
+    741,
+    790,
+    791,
+    793,
+    794,
+    795,
+    800,
+    819,
+    829,
+    830,
+    844,
+    845
+  ]
+}
\ No newline at end of file
diff --git a/fr/1120ms/preprocessor.mlmodelc/analytics/coremldata.bin b/fr/1120ms/preprocessor.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..e80c07518c7261d08beb9bc52cd276894a4bd14c
--- /dev/null
+++ b/fr/1120ms/preprocessor.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:12264cf4116cb366a0a4cc606ffba954ad2f9494d25776e76112c294b847ffb7
+size 243
diff --git a/fr/1120ms/preprocessor.mlmodelc/coremldata.bin b/fr/1120ms/preprocessor.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..22424e803aac20ad868e8f2d398295f9b0f919d3
--- /dev/null
+++ b/fr/1120ms/preprocessor.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66115dc0006e9aac8b4b2340806691cb4083df62b5757a8e5861b3e30b6065c6
+size 371
diff --git a/fr/1120ms/preprocessor.mlmodelc/model.mil b/fr/1120ms/preprocessor.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..0b8261362f9cbf465b530a0d2d0ee9a2b2f462cd
--- /dev/null
+++ b/fr/1120ms/preprocessor.mlmodelc/model.mil
@@ -0,0 +1,122 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp32, [1, ?]> audio, tensor<int32, [1]> audio_length) [FlexibleShapeInformation = tuple<tuple<string, dict<string, tensor<int32, [?]>>>, tuple<string, dict<string, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"audio", [1, 1]}}), ("RangeDims", {{"audio", [[1, 1], [1, 480000]]}})))] {
+            int32 var_9 = const()[name = string("op_9"), val = int32(1)];
+            int32 var_10 = const()[name = string("op_10"), val = int32(160)];
+            int32 var_12 = const()[name = string("op_12"), val = int32(0)];
+            int32 var_33 = const()[name = string("op_33"), val = int32(512)];
+            tensor<int32, [1]> var_34 = add(x = audio_length, y = var_33)[name = string("op_34")];
+            int32 var_35 = const()[name = string("op_35"), val = int32(512)];
+            tensor<int32, [1]> var_36 = sub(x = var_34, y = var_35)[name = string("op_36")];
+            tensor<int32, [1]> floor_div_0 = floor_div(x = var_36, y = var_10)[name = string("floor_div_0")];
+            tensor<bool, [1]> var_39 = equal(x = audio_length, y = var_12)[name = string("op_39")];
+            tensor<int32, [1]> var_40 = const()[name = string("op_40"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> mel_length = select(a = var_40, b = floor_div_0, cond = var_39)[name = string("seq_len")];
+            string audio_to_fp16_dtype_0 = const()[name = string("audio_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, ?]> audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = string("cast_14")];
+            tensor<int32, [2]> var_42_shape_cast_fp16 = shape(x = audio_to_fp16)[name = string("op_42_shape_cast_fp16")];
+            int32 gather_0_axis_0 = const()[name = string("gather_0_axis_0"), val = int32(0)];
+            int32 gather_0_batch_dims_0 = const()[name = string("gather_0_batch_dims_0"), val = int32(0)];
+            bool gather_0_validate_indices_0 = const()[name = string("gather_0_validate_indices_0"), val = bool(false)];
+            string var_42_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_42_shape_cast_fp16_to_int16_dtype_0"), val = string("int16")];
+            uint16 select_0_to_uint16 = const()[name = string("select_0_to_uint16"), val = uint16(1)];
+            tensor<int16, [2]> var_42_shape_cast_fp16_to_int16 = cast(dtype = var_42_shape_cast_fp16_to_int16_dtype_0, x = var_42_shape_cast_fp16)[name = string("cast_13")];
+            int16 gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_42_shape_cast_fp16_to_int16)[name = string("gather_0_cast_uint16")];
+            string gather_0_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_0_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            int32 const_0 = const()[name = string("const_0"), val = int32(0)];
+            int32 const_1 = const()[name = string("const_1"), val = int32(1)];
+            int32 gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = string("cast_12")];
+            tensor<int32, [?]> var_43 = range_1d(end = gather_0_cast_uint16_to_int32, start = const_0, step = const_1)[name = string("op_43")];
+            tensor<int32, [1]> var_44_axes_0 = const()[name = string("op_44_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> var_44 = expand_dims(axes = var_44_axes_0, x = var_43)[name = string("op_44")];
+            tensor<int32, [1]> var_45_axes_0 = const()[name = string("op_45_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_45 = expand_dims(axes = var_45_axes_0, x = audio_length)[name = string("op_45")];
+            tensor<bool, [1, ?]> timemask = less(x = var_44, y = var_45)[name = string("timemask")];
+            tensor<int32, [2]> var_48_begin_0 = const()[name = string("op_48_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_48_end_0 = const()[name = string("op_48_end_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<bool, [2]> var_48_end_mask_0 = const()[name = string("op_48_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<bool, [2]> var_48_squeeze_mask_0 = const()[name = string("op_48_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
+            tensor<fp16, [1]> var_48_cast_fp16 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, squeeze_mask = var_48_squeeze_mask_0, x = audio_to_fp16)[name = string("op_48_cast_fp16")];
+            tensor<int32, [1]> var_49_axes_0 = const()[name = string("op_49_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1]> var_49_cast_fp16 = expand_dims(axes = var_49_axes_0, x = var_48_cast_fp16)[name = string("op_49_cast_fp16")];
+            tensor<int32, [2]> var_51_begin_0 = const()[name = string("op_51_begin_0"), val = tensor<int32, [2]>([0, 1])];
+            tensor<int32, [2]> var_51_end_0 = const()[name = string("op_51_end_0"), val = tensor<int32, [2]>([1, 0])];
+            tensor<bool, [2]> var_51_end_mask_0 = const()[name = string("op_51_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<fp16, [1, ?]> var_51_cast_fp16 = slice_by_index(begin = var_51_begin_0, end = var_51_end_0, end_mask = var_51_end_mask_0, x = audio_to_fp16)[name = string("op_51_cast_fp16")];
+            tensor<int32, [2]> var_53_begin_0 = const()[name = string("op_53_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_53_end_0 = const()[name = string("op_53_end_0"), val = tensor<int32, [2]>([1, -1])];
+            tensor<bool, [2]> var_53_end_mask_0 = const()[name = string("op_53_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp16, [1, ?]> var_53_cast_fp16 = slice_by_index(begin = var_53_begin_0, end = var_53_end_0, end_mask = var_53_end_mask_0, x = audio_to_fp16)[name = string("op_53_cast_fp16")];
+            fp16 var_54_to_fp16 = const()[name = string("op_54_to_fp16"), val = fp16(0x1.f0cp-1)];
+            tensor<fp16, [1, ?]> var_55_cast_fp16 = mul(x = var_53_cast_fp16, y = var_54_to_fp16)[name = string("op_55_cast_fp16")];
+            tensor<fp16, [1, ?]> var_56_cast_fp16 = sub(x = var_51_cast_fp16, y = var_55_cast_fp16)[name = string("op_56_cast_fp16")];
+            bool x_3_interleave_0 = const()[name = string("x_3_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, ?]> x_3_cast_fp16 = concat(axis = var_9, interleave = x_3_interleave_0, values = (var_49_cast_fp16, var_56_cast_fp16))[name = string("x_3_cast_fp16")];
+            tensor<bool, [1, ?]> var_59 = logical_not(x = timemask)[name = string("op_59")];
+            fp16 var_16_to_fp16 = const()[name = string("op_16_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, ?]> input_1_cast_fp16 = select(a = var_16_to_fp16, b = x_3_cast_fp16, cond = var_59)[name = string("input_1_cast_fp16")];
+            tensor<int32, [3]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, ?]> input_3_cast_fp16 = reshape(shape = concat_1x, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")];
+            tensor<int32, [6]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 256, 256])];
+            string input_5_mode_0 = const()[name = string("input_5_mode_0"), val = string("constant")];
+            fp16 const_3_to_fp16 = const()[name = string("const_3_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, ?]> input_5_cast_fp16 = pad(constant_val = const_3_to_fp16, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<int32, [2]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [2]>([1, -1])];
+            tensor<fp16, [1, ?]> input_cast_fp16 = reshape(shape = concat_2x, x = input_5_cast_fp16)[name = string("input_cast_fp16")];
+            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([160])];
+            tensor<int32, [1]> expand_dims_4_axes_0 = const()[name = string("expand_dims_4_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, ?]> expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = input_cast_fp16)[name = string("expand_dims_4_cast_fp16")];
+            string conv_0_pad_type_0 = const()[name = string("conv_0_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> conv_0_pad_0 = const()[name = string("conv_0_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_0_dilations_0 = const()[name = string("conv_0_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 conv_0_groups_0 = const()[name = string("conv_0_groups_0"), val = int32(1)];
+            tensor<fp16, [257, 1, 512]> expand_dims_1_to_fp16 = const()[name = string("expand_dims_1_to_fp16"), val = tensor<fp16, [257, 1, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 257, ?]> conv_0_cast_fp16 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_0_cast_fp16")];
+            string conv_1_pad_type_0 = const()[name = string("conv_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> conv_1_pad_0 = const()[name = string("conv_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_1_dilations_0 = const()[name = string("conv_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 conv_1_groups_0 = const()[name = string("conv_1_groups_0"), val = int32(1)];
+            tensor<fp16, [257, 1, 512]> expand_dims_2_to_fp16 = const()[name = string("expand_dims_2_to_fp16"), val = tensor<fp16, [257, 1, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263296)))];
+            tensor<fp16, [1, 257, ?]> conv_1_cast_fp16 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_1_cast_fp16")];
+            int32 stack_0_axis_0 = const()[name = string("stack_0_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 257, ?, 2]> stack_0_cast_fp16 = stack(axis = stack_0_axis_0, values = (conv_0_cast_fp16, conv_1_cast_fp16))[name = string("stack_0_cast_fp16")];
+            fp16 var_19_promoted_to_fp16 = const()[name = string("op_19_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1, 257, ?, 2]> var_74_cast_fp16 = pow(x = stack_0_cast_fp16, y = var_19_promoted_to_fp16)[name = string("op_74_cast_fp16")];
+            tensor<int32, [1]> var_76_axes_0 = const()[name = string("op_76_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_76_keep_dims_0 = const()[name = string("op_76_keep_dims_0"), val = bool(false)];
+            tensor<fp16, [1, 257, ?]> var_76_cast_fp16 = reduce_sum(axes = var_76_axes_0, keep_dims = var_76_keep_dims_0, x = var_74_cast_fp16)[name = string("op_76_cast_fp16")];
+            tensor<fp16, [1, 257, ?]> x_11_cast_fp16 = identity(x = var_76_cast_fp16)[name = string("x_11_cast_fp16")];
+            bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)];
+            bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 128, 257]> const_4_to_fp16 = const()[name = string("const_4_to_fp16"), val = tensor<fp16, [1, 128, 257]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526528)))];
+            tensor<fp16, [1, 128, ?]> x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = const_4_to_fp16, y = x_11_cast_fp16)[name = string("x_13_cast_fp16")];
+            fp16 var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 128, ?]> var_84_cast_fp16 = add(x = x_13_cast_fp16, y = var_83_to_fp16)[name = string("op_84_cast_fp16")];
+            fp32 x_epsilon_0 = const()[name = string("x_epsilon_0"), val = fp32(0x1p-149)];
+            tensor<fp16, [1, 128, ?]> x_cast_fp16 = log(epsilon = x_epsilon_0, x = var_84_cast_fp16)[name = string("x_cast_fp16")];
+            tensor<int32, [3]> var_86_shape_cast_fp16 = shape(x = x_cast_fp16)[name = string("op_86_shape_cast_fp16")];
+            int32 gather_5_axis_0 = const()[name = string("gather_5_axis_0"), val = int32(0)];
+            int32 gather_5_batch_dims_0 = const()[name = string("gather_5_batch_dims_0"), val = int32(0)];
+            bool gather_5_validate_indices_0 = const()[name = string("gather_5_validate_indices_0"), val = bool(false)];
+            string var_86_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_86_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
+            uint16 select_5_to_uint16 = const()[name = string("select_5_to_uint16"), val = uint16(2)];
+            tensor<uint16, [3]> var_86_shape_cast_fp16_to_uint16 = cast(dtype = var_86_shape_cast_fp16_to_uint16_dtype_0, x = var_86_shape_cast_fp16)[name = string("cast_11")];
+            uint16 gather_5_cast_uint16 = gather(axis = gather_5_axis_0, batch_dims = gather_5_batch_dims_0, indices = select_5_to_uint16, validate_indices = gather_5_validate_indices_0, x = var_86_shape_cast_fp16_to_uint16)[name = string("gather_5_cast_uint16")];
+            string gather_5_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_5_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            int32 const_5 = const()[name = string("const_5"), val = int32(0)];
+            int32 const_6 = const()[name = string("const_6"), val = int32(1)];
+            int32 gather_5_cast_uint16_to_int32 = cast(dtype = gather_5_cast_uint16_to_int32_dtype_0, x = gather_5_cast_uint16)[name = string("cast_10")];
+            tensor<int32, [?]> mask_1 = range_1d(end = gather_5_cast_uint16_to_int32, start = const_5, step = const_6)[name = string("mask_1")];
+            tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = string("expand_dims_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = string("expand_dims_0")];
+            tensor<int32, [1]> var_91_axes_0 = const()[name = string("op_91_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_91 = expand_dims(axes = var_91_axes_0, x = mel_length)[name = string("op_91")];
+            tensor<bool, [1, ?]> mask = greater_equal(x = expand_dims_0, y = var_91)[name = string("mask")];
+            tensor<int32, [1]> var_93_axes_0 = const()[name = string("op_93_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, ?]> var_93 = expand_dims(axes = var_93_axes_0, x = mask)[name = string("op_93")];
+            tensor<fp16, [1, 128, ?]> processed_signal_cast_fp16 = select(a = var_16_to_fp16, b = x_cast_fp16, cond = var_93)[name = string("processed_signal_cast_fp16")];
+            string processed_signal_cast_fp16_to_fp32_dtype_0 = const()[name = string("processed_signal_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 128, ?]> mel = cast(dtype = processed_signal_cast_fp16_to_fp32_dtype_0, x = processed_signal_cast_fp16)[name = string("cast_9")];
+        } -> (mel, mel_length);
+}
\ No newline at end of file
diff --git a/fr/1120ms/preprocessor.mlmodelc/weights/weight.bin b/fr/1120ms/preprocessor.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411
--- /dev/null
+++ b/fr/1120ms/preprocessor.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3
+size 592384
diff --git a/fr/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..d1f9d01bb6fca81fc2a3212d0fd405343a5c5971
--- /dev/null
+++ b/fr/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:daad3e5dc1a13a1dc66997f9c1c1280b68646d7ccd7b6359fa8c361d63b5c09b
+size 15878
diff --git a/fr/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411
--- /dev/null
+++ b/fr/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3
+size 592384
diff --git a/fr/1120ms/preprocessor.mlpackage/Manifest.json b/fr/1120ms/preprocessor.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..9722979403b2b2f0a821e990b4205cca52694033
--- /dev/null
+++ b/fr/1120ms/preprocessor.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "11EFF503-1D75-4BC3-9480-44E97E36229D": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "A4EAB1E5-DDC2-4445-9D00-C29BCDE18F72": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "A4EAB1E5-DDC2-4445-9D00-C29BCDE18F72"
+}
diff --git a/fr/1120ms/tokenizer.json b/fr/1120ms/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..53c1afe82c2a72808c78fffb846345d590c453fd
--- /dev/null
+++ b/fr/1120ms/tokenizer.json
@@ -0,0 +1,851 @@
+{
+  "0": "<unk>",
+  "1": "<bg-BG>",
+  "2": "▁",
+  "3": ".",
+  "4": ",",
+  "5": "e",
+  "6": "t",
+  "7": "a",
+  "8": "s",
+  "9": "o",
+  "10": "i",
+  "11": "r",
+  "12": "l",
+  "13": "u",
+  "14": "d",
+  "15": "c",
+  "16": "h",
+  "17": "m",
+  "18": "p",
+  "19": "n",
+  "20": "g",
+  "21": "f",
+  "22": "en",
+  "23": "in",
+  "24": "on",
+  "25": "y",
+  "26": "er",
+  "27": "an",
+  "28": "w",
+  "29": "0",
+  "30": "b",
+  "31": "v",
+  "32": "2",
+  "33": "1",
+  "34": "▁the",
+  "35": "k",
+  "36": "3",
+  "37": "5",
+  "38": "q",
+  "39": "I",
+  "40": "é",
+  "41": "4",
+  "42": "z",
+  "43": "A",
+  "44": "6",
+  "45": "j",
+  "46": "E",
+  "47": "7",
+  "48": "T",
+  "49": "8",
+  "50": "9",
+  "51": "S",
+  "52": "x",
+  "53": "C",
+  "54": "M",
+  "55": "P",
+  "56": "D",
+  "57": "B",
+  "58": "U",
+  "59": "N",
+  "60": "O",
+  "61": "G",
+  "62": "à",
+  "63": "L",
+  "64": "è",
+  "65": "V",
+  "66": "R",
+  "67": "ó",
+  "68": "J",
+  "69": "K",
+  "70": "ê",
+  "71": "ç",
+  "72": "<cs-CZ>",
+  "73": "st",
+  "74": "ch",
+  "75": "le",
+  "76": "li",
+  "77": "▁po",
+  "78": "no",
+  "79": "to",
+  "80": "me",
+  "81": "te",
+  "82": "ho",
+  "83": "▁pro",
+  "84": "ro",
+  "85": "▁na",
+  "86": "ce",
+  "87": "la",
+  "88": "ni",
+  "89": "ra",
+  "90": "ti",
+  "91": "lo",
+  "92": "ko",
+  "93": "po",
+  "94": "je",
+  "95": "de",
+  "96": "na",
+  "97": "mi",
+  "98": "ci",
+  "99": "▁by",
+  "100": "ve",
+  "101": "▁za",
+  "102": "▁A",
+  "103": "re",
+  "104": "ou",
+  "105": "vo",
+  "106": "né",
+  "107": "va",
+  "108": "mo",
+  "109": "ze",
+  "110": "ne",
+  "111": "ka",
+  "112": "▁jak",
+  "113": "ny",
+  "114": "vi",
+  "115": "ent",
+  "116": "▁pan",
+  "117": "Z",
+  "118": "▁To",
+  "119": "▁Je",
+  "120": "ă",
+  "121": "X",
+  "122": "<da-DK>",
+  "123": "▁for",
+  "124": "▁at",
+  "125": "et",
+  "126": "▁vi",
+  "127": "al",
+  "128": "▁de",
+  "129": "▁der",
+  "130": "or",
+  "131": "om",
+  "132": "and",
+  "133": "at",
+  "134": "▁af",
+  "135": "ge",
+  "136": "ar",
+  "137": "is",
+  "138": "▁be",
+  "139": "un",
+  "140": "lig",
+  "141": "▁man",
+  "142": "ig",
+  "143": "▁som",
+  "144": "el",
+  "145": "ag",
+  "146": "erne",
+  "147": "▁den",
+  "148": "ste",
+  "149": "id",
+  "150": "ske",
+  "151": "iv",
+  "152": "▁her",
+  "153": "ion",
+  "154": "am",
+  "155": "ur",
+  "156": "for",
+  "157": "▁pr",
+  "158": "▁sig",
+  "159": "▁ind",
+  "160": "ende",
+  "161": "▁Vi",
+  "162": "ation",
+  "163": "mme",
+  "164": "tage",
+  "165": "▁op",
+  "166": "▁fra",
+  "167": "▁alle",
+  "168": "▁Men",
+  "169": "▁var",
+  "170": "<de-DE>",
+  "171": "▁die",
+  "172": "sch",
+  "173": "it",
+  "174": "gen",
+  "175": "▁W",
+  "176": "▁B",
+  "177": "▁E",
+  "178": "▁F",
+  "179": "ll",
+  "180": "▁es",
+  "181": "▁K",
+  "182": "ie",
+  "183": "au",
+  "184": "ich",
+  "185": "ck",
+  "186": "ten",
+  "187": "mal",
+  "188": "ein",
+  "189": "▁T",
+  "190": "▁mit",
+  "191": "hr",
+  "192": "ter",
+  "193": "tz",
+  "194": "▁G",
+  "195": "ben",
+  "196": "um",
+  "197": "us",
+  "198": "cht",
+  "199": "il",
+  "200": "ut",
+  "201": "▁ver",
+  "202": "▁hier",
+  "203": "▁von",
+  "204": "ri",
+  "205": "ach",
+  "206": "ol",
+  "207": "▁Da",
+  "208": "sp",
+  "209": "ell",
+  "210": "▁ja",
+  "211": "uch",
+  "212": "wi",
+  "213": "rei",
+  "214": "▁Ge",
+  "215": "und",
+  "216": "isch",
+  "217": "▁sie",
+  "218": "▁Ja",
+  "219": "▁du",
+  "220": "▁dem",
+  "221": "▁Sch",
+  "222": "▁Ma",
+  "223": "▁De",
+  "224": "▁Le",
+  "225": "▁In",
+  "226": "▁Re",
+  "227": "▁sein",
+  "228": "▁Mi",
+  "229": "▁Ha",
+  "230": "<el-GR>",
+  "231": "<et-EE>",
+  "232": "ma",
+  "233": "ta",
+  "234": "se",
+  "235": "da",
+  "236": "si",
+  "237": "ks",
+  "238": "ga",
+  "239": "he",
+  "240": "mu",
+  "241": "tu",
+  "242": "ha",
+  "243": "ja",
+  "244": "gi",
+  "245": "nd",
+  "246": "oo",
+  "247": "gu",
+  "248": "ju",
+  "249": "est",
+  "250": "▁pa",
+  "251": "tud",
+  "252": "▁pea",
+  "253": "sta",
+  "254": "di",
+  "255": "mise",
+  "256": "▁su",
+  "257": "ide",
+  "258": "val",
+  "259": "▁Me",
+  "260": "ment",
+  "261": "▁meil",
+  "262": "Q",
+  "263": "<fi-FI>",
+  "264": "ssa",
+  "265": "lla",
+  "266": "ty",
+  "267": "ki",
+  "268": "pa",
+  "269": "lle",
+  "270": "lu",
+  "271": "tta",
+  "272": "isi",
+  "273": "ise",
+  "274": "tte",
+  "275": "ista",
+  "276": "inen",
+  "277": "llis",
+  "278": "vu",
+  "279": "▁voi",
+  "280": "iden",
+  "281": "ë",
+  "282": "<fr-FR>",
+  "283": "▁est",
+  "284": "▁c",
+  "285": "▁que",
+  "286": "▁à",
+  "287": "es",
+  "288": "▁un",
+  "289": "▁pas",
+  "290": "▁les",
+  "291": "▁qui",
+  "292": "▁il",
+  "293": "▁vous",
+  "294": "▁des",
+  "295": "▁qu",
+  "296": "▁pour",
+  "297": "▁par",
+  "298": "▁ça",
+  "299": "▁une",
+  "300": "ant",
+  "301": "ais",
+  "302": "ez",
+  "303": "▁dans",
+  "304": "▁C",
+  "305": "tre",
+  "306": "ir",
+  "307": "elle",
+  "308": "eur",
+  "309": "▁sur",
+  "310": "▁con",
+  "311": "▁Et",
+  "312": "ement",
+  "313": "tion",
+  "314": "té",
+  "315": "▁tout",
+  "316": "mp",
+  "317": "ique",
+  "318": "▁plus",
+  "319": "eux",
+  "320": "▁dé",
+  "321": "▁fait",
+  "322": "qu",
+  "323": "▁comme",
+  "324": "ac",
+  "325": "ait",
+  "326": "che",
+  "327": "▁mais",
+  "328": "que",
+  "329": "ul",
+  "330": "▁avec",
+  "331": "▁bien",
+  "332": "age",
+  "333": "▁mon",
+  "334": "end",
+  "335": "▁faire",
+  "336": "▁être",
+  "337": "ver",
+  "338": "▁peu",
+  "339": "▁même",
+  "340": "tra",
+  "341": "cha",
+  "342": "▁nous",
+  "343": "▁donc",
+  "344": "▁sont",
+  "345": "▁moi",
+  "346": "ille",
+  "347": "ff",
+  "348": "▁ex",
+  "349": "▁Il",
+  "350": "▁très",
+  "351": "▁cette",
+  "352": "im",
+  "353": "ité",
+  "354": "▁dire",
+  "355": "▁peut",
+  "356": "ance",
+  "357": "aire",
+  "358": "mé",
+  "359": "▁app",
+  "360": "▁aussi",
+  "361": "▁petit",
+  "362": "aux",
+  "363": "▁parce",
+  "364": "onne",
+  "365": "mb",
+  "366": "man",
+  "367": "▁On",
+  "368": "▁quand",
+  "369": "▁autre",
+  "370": "ô",
+  "371": "▁chose",
+  "372": "▁puis",
+  "373": "▁était",
+  "374": "ndre",
+  "375": "port",
+  "376": "▁vraiment",
+  "377": "ence",
+  "378": "▁Mais",
+  "379": "î",
+  "380": "▁avoir",
+  "381": "form",
+  "382": "▁faut",
+  "383": "ign",
+  "384": "▁où",
+  "385": "près",
+  "386": "▁beaucoup",
+  "387": "ture",
+  "388": "û",
+  "389": "â",
+  "390": "<hu-HU>",
+  "391": "▁az",
+  "392": "ok",
+  "393": "gy",
+  "394": "ek",
+  "395": "és",
+  "396": "em",
+  "397": "▁is",
+  "398": "os",
+  "399": "ban",
+  "400": "ik",
+  "401": "cs",
+  "402": "él",
+  "403": "ér",
+  "404": "nek",
+  "405": "ség",
+  "406": "én",
+  "407": "hat",
+  "408": "▁ha",
+  "409": "ép",
+  "410": "vé",
+  "411": "leg",
+  "412": "▁ami",
+  "413": "É",
+  "414": "rend",
+  "415": "<hr-HR>",
+  "416": "▁bi",
+  "417": "▁sa",
+  "418": "ru",
+  "419": "go",
+  "420": "sti",
+  "421": "▁pri",
+  "422": "ima",
+  "423": "nu",
+  "424": "▁pre",
+  "425": "nja",
+  "426": "zi",
+  "427": "vr",
+  "428": "ca",
+  "429": "ba",
+  "430": "<it-IT>",
+  "431": "▁di",
+  "432": "▁che",
+  "433": "co",
+  "434": "▁per",
+  "435": "▁non",
+  "436": "do",
+  "437": "gli",
+  "438": "so",
+  "439": "amo",
+  "440": "sa",
+  "441": "ndo",
+  "442": "fi",
+  "443": "pi",
+  "444": "nti",
+  "445": "tto",
+  "446": "tro",
+  "447": "▁fa",
+  "448": "chi",
+  "449": "▁qua",
+  "450": "bi",
+  "451": "▁del",
+  "452": "pe",
+  "453": "ssi",
+  "454": "▁sono",
+  "455": "nte",
+  "456": "sso",
+  "457": "nta",
+  "458": "cu",
+  "459": "ggi",
+  "460": "sto",
+  "461": "gra",
+  "462": "spe",
+  "463": "cco",
+  "464": "nde",
+  "465": "fe",
+  "466": "pu",
+  "467": "zza",
+  "468": "sci",
+  "469": "▁poi",
+  "470": "sco",
+  "471": "stra",
+  "472": "▁quel",
+  "473": "qui",
+  "474": "sse",
+  "475": "▁inter",
+  "476": "sce",
+  "477": "▁com",
+  "478": "▁comp",
+  "479": "▁prima",
+  "480": "▁imp",
+  "481": "▁cui",
+  "482": "scri",
+  "483": "<lt-LT>",
+  "484": "ai",
+  "485": "▁ir",
+  "486": "as",
+  "487": "▁tai",
+  "488": "uo",
+  "489": "tin",
+  "490": "▁vis",
+  "491": "ly",
+  "492": "gal",
+  "493": "jo",
+  "494": "tar",
+  "495": "yb",
+  "496": "▁Ir",
+  "497": "▁Tai",
+  "498": "▁nu",
+  "499": "▁mes",
+  "500": "▁pra",
+  "501": "<lv-LV>",
+  "502": "▁pie",
+  "503": "ies",
+  "504": "ot",
+  "505": "▁Un",
+  "506": "▁Ta",
+  "507": "▁lie",
+  "508": "dar",
+  "509": "sim",
+  "510": "dien",
+  "511": "gan",
+  "512": "▁ap",
+  "513": "▁nav",
+  "514": "▁cit",
+  "515": "▁Ne",
+  "516": "zin",
+  "517": "reiz",
+  "518": "▁20",
+  "519": "<nl-NL>",
+  "520": "▁dat",
+  "521": "▁we",
+  "522": "▁En",
+  "523": "▁dan",
+  "524": "▁met",
+  "525": "▁wat",
+  "526": "der",
+  "527": "ui",
+  "528": "den",
+  "529": "op",
+  "530": "oor",
+  "531": "▁of",
+  "532": "ven",
+  "533": "acht",
+  "534": "ond",
+  "535": "▁dit",
+  "536": "▁wil",
+  "537": "▁We",
+  "538": "vol",
+  "539": "llen",
+  "540": "▁Dan",
+  "541": "▁Nou",
+  "542": "<pl-PL>",
+  "543": "nie",
+  "544": "wa",
+  "545": "my",
+  "546": "cie",
+  "547": "nia",
+  "548": "wo",
+  "549": "rze",
+  "550": "by",
+  "551": "za",
+  "552": "dy",
+  "553": "ry",
+  "554": "mie",
+  "555": "pie",
+  "556": "uje",
+  "557": "▁bo",
+  "558": "bie",
+  "559": "sy",
+  "560": "▁Po",
+  "561": "ski",
+  "562": "▁No",
+  "563": "▁Na",
+  "564": "▁musi",
+  "565": "<pt-BR>",
+  "566": "pt",
+  "567": "-",
+  "568": "▁eu",
+  "569": "▁para",
+  "570": "▁pe",
+  "571": "▁tem",
+  "572": "▁O",
+  "573": "pre",
+  "574": "ria",
+  "575": "▁fo",
+  "576": "mos",
+  "577": "bo",
+  "578": "▁por",
+  "579": "▁Eu",
+  "580": "be",
+  "581": "▁esse",
+  "582": "ente",
+  "583": "▁mas",
+  "584": "qua",
+  "585": "fica",
+  "586": "▁vou",
+  "587": "▁Se",
+  "588": "▁fala",
+  "589": "...",
+  "590": "ço",
+  "591": "▁Por",
+  "592": "▁Co",
+  "593": "iza",
+  "594": "<ro-RO>",
+  "595": "▁mai",
+  "596": "sc",
+  "597": "are",
+  "598": "▁din",
+  "599": "rea",
+  "600": "du",
+  "601": "▁M",
+  "602": "▁fac",
+  "603": "lor",
+  "604": "▁mult",
+  "605": "per",
+  "606": "cur",
+  "607": "tor",
+  "608": "inte",
+  "609": "▁sau",
+  "610": "tat",
+  "611": "ori",
+  "612": "▁prim",
+  "613": "▁lui",
+  "614": "▁sub",
+  "615": "▁prin",
+  "616": "▁alt",
+  "617": "stru",
+  "618": "fer",
+  "619": "<ru-RU>",
+  "620": "<sk-SK>",
+  "621": "sk",
+  "622": "▁aj",
+  "623": "ob",
+  "624": "ali",
+  "625": "nov",
+  "626": "rov",
+  "627": "rob",
+  "628": "▁spo",
+  "629": "<sl-SL>",
+  "630": "sl",
+  "631": "udi",
+  "632": "del",
+  "633": "▁sem",
+  "634": "▁samo",
+  "635": "▁let",
+  "636": "▁Pre",
+  "637": "▁prot",
+  "638": "▁internet",
+  "639": "▁film",
+  "640": "<sv-SE>",
+  "641": "▁att",
+  "642": "▁inte",
+  "643": "▁av",
+  "644": "all",
+  "645": "era",
+  "646": "pp",
+  "647": "sam",
+  "648": "▁vill",
+  "649": "erna",
+  "650": "ande",
+  "651": "ade",
+  "652": "▁hur",
+  "653": "bil",
+  "654": "▁min",
+  "655": "lev",
+  "656": "land",
+  "657": "text",
+  "658": "person",
+  "659": "<uk-UA>",
+  "660": "<ar-AR>",
+  "661": "'",
+  "662": "▁H",
+  "663": "▁D",
+  "664": "aus",
+  "665": "▁N",
+  "666": "▁Be",
+  "667": "mm",
+  "668": "ab",
+  "669": "▁Er",
+  "670": "ssen",
+  "671": "hl",
+  "672": "rie",
+  "673": "▁An",
+  "674": "rau",
+  "675": "▁So",
+  "676": "▁Aus",
+  "677": "<en-GB>",
+  "678": "▁can",
+  "679": "ed",
+  "680": "ay",
+  "681": "th",
+  "682": "ic",
+  "683": "hi",
+  "684": "▁not",
+  "685": "ight",
+  "686": "ex",
+  "687": "▁four",
+  "688": "▁great",
+  "689": "▁six",
+  "690": "ill",
+  "691": "▁but",
+  "692": "▁got",
+  "693": "▁don",
+  "694": "▁out",
+  "695": "▁from",
+  "696": "▁fine",
+  "697": "able",
+  "698": "ever",
+  "699": "▁month",
+  "700": "▁question",
+  "701": "▁possible",
+  "702": "▁moment",
+  "703": "▁change",
+  "704": "<en-US>",
+  "705": "ow",
+  "706": "ive",
+  "707": "ate",
+  "708": "ad",
+  "709": "ect",
+  "710": "ng",
+  "711": "ther",
+  "712": "act",
+  "713": "ist",
+  "714": "▁part",
+  "715": "side",
+  "716": "\"",
+  "717": "<es-ES>",
+  "718": "ción",
+  "719": "res",
+  "720": "▁La",
+  "721": "dos",
+  "722": "▁las",
+  "723": "men",
+  "724": "par",
+  "725": "rio",
+  "726": "cho",
+  "727": "ble",
+  "728": "▁Ca",
+  "729": "▁Su",
+  "730": "▁son",
+  "731": "▁Con",
+  "732": "ones",
+  "733": "ecto",
+  "734": "▁Com",
+  "735": "<es-US>",
+  "736": "cia",
+  "737": "▁Y",
+  "738": "ron",
+  "739": "les",
+  "740": "bu",
+  "741": "<fr-CA>",
+  "742": "ré",
+  "743": "ée",
+  "744": "▁Les",
+  "745": "our",
+  "746": "▁Ce",
+  "747": "com",
+  "748": "▁Elle",
+  "749": "▁Cet",
+  "750": "ux",
+  "751": "ale",
+  "752": "ction",
+  "753": "▁pré",
+  "754": "▁deux",
+  "755": "if",
+  "756": "lé",
+  "757": "ère",
+  "758": "ière",
+  "759": "iste",
+  "760": "▁parti",
+  "761": "▁été",
+  "762": "avec",
+  "763": "jour",
+  "764": "app",
+  "765": "cul",
+  "766": "▁égale",
+  "767": "aine",
+  "768": "gue",
+  "769": "▁nombre",
+  "770": "▁étai",
+  "771": "tout",
+  "772": "▁grand",
+  "773": "▁commun",
+  "774": "Une",
+  "775": "œ",
+  "776": "ï",
+  "777": "À",
+  "778": "Â",
+  "779": "av",
+  "780": "ouv",
+  "781": "êt",
+  "782": "ois",
+  "783": "pri",
+  "784": "voir",
+  "785": "sion",
+  "786": "ix",
+  "787": "était",
+  "788": "ard",
+  "789": "aient",
+  "790": "<he-IL>",
+  "791": "<hi-IN>",
+  "792": "!",
+  "793": "<ja-JP>",
+  "794": "<ko-KR>",
+  "795": "<nb-NO>",
+  "796": "▁ble",
+  "797": "opp",
+  "798": "▁Han",
+  "799": "▁han",
+  "800": "<nn-NO>",
+  "801": "eg",
+  "802": "kk",
+  "803": "dig",
+  "804": "ord",
+  "805": "▁tru",
+  "806": "▁sei",
+  "807": "ller",
+  "808": "▁fle",
+  "809": "car",
+  "810": "▁foi",
+  "811": "ito",
+  "812": "ram",
+  "813": "fa",
+  "814": "▁meu",
+  "815": "nça",
+  "816": "▁mil",
+  "817": "▁passa",
+  "818": "▁casa",
+  "819": "<pt-PT>",
+  "820": "das",
+  "821": "▁Pa",
+  "822": "forma",
+  "823": "mar",
+  "824": "▁entre",
+  "825": "fun",
+  "826": "▁grande",
+  "827": "▁Este",
+  "828": "var",
+  "829": "<th-TH>",
+  "830": "<tr-TR>",
+  "831": "ş",
+  "832": "ğ",
+  "833": "ya",
+  "834": "▁ve",
+  "835": "lar",
+  "836": "ler",
+  "837": "ye",
+  "838": "▁bu",
+  "839": "lan",
+  "840": "ara",
+  "841": "▁Bu",
+  "842": "inde",
+  "843": "yo",
+  "844": "<zh-CN>",
+  "845": "<vi-VN>",
+  "846": "▁t",
+  "847": "nh",
+  "848": "<blank>"
+}
\ No newline at end of file
diff --git a/fr/2240ms/decoder.mlmodelc/analytics/coremldata.bin b/fr/2240ms/decoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..65b800e186dcb664ec3d46edb11fa43215dce1f7
--- /dev/null
+++ b/fr/2240ms/decoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d3becf0abd7a7efd30577f33f4144cfcb8c08155fdd62903eb1a45d7b02b87c
+size 243
diff --git a/fr/2240ms/decoder.mlmodelc/coremldata.bin b/fr/2240ms/decoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..a921778bfd0fef1f29bd148ddcb6c74c6116663e
--- /dev/null
+++ b/fr/2240ms/decoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d6b7630abf0c1b447913bf9d1ed5935d625bcb70e978c607820b2ff35ee8f07f
+size 433
diff --git a/fr/2240ms/decoder.mlmodelc/model.mil b/fr/2240ms/decoder.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..0de41956c1a2cac40f0629b26d6b60b23d5fe6b3
--- /dev/null
+++ b/fr/2240ms/decoder.mlmodelc/model.mil
@@ -0,0 +1,64 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            int32 y_axis_0 = const()[name = string("y_axis_0"), val = int32(0)];
+            int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)];
+            bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [849, 640]> module_prediction_embed_weight_to_fp16 = const()[name = string("module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_8")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)];
+            int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)];
+            string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")];
+            int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)];
+            int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)];
+            string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            string input_lstm_layer_0_direction_0 = const()[name = string("input_lstm_layer_0_direction_0"), val = string("forward")];
+            bool input_lstm_layer_0_output_sequence_0 = const()[name = string("input_lstm_layer_0_output_sequence_0"), val = bool(true)];
+            string input_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")];
+            string input_lstm_layer_0_cell_activation_0 = const()[name = string("input_lstm_layer_0_cell_activation_0"), val = string("tanh")];
+            string input_lstm_layer_0_activation_0 = const()[name = string("input_lstm_layer_0_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1086848)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4363712)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7640576)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = string("transpose_2")];
+            tensor<fp16, [1, 1, 640]> input_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_lstm_layer_0_cast_fp16_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_lstm_layer_0_cell_activation_0, direction = input_lstm_layer_0_direction_0, initial_c = input_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_lstm_c0_squeeze_cast_fp16")];
+            string input_direction_0 = const()[name = string("input_direction_0"), val = string("forward")];
+            bool input_output_sequence_0 = const()[name = string("input_output_sequence_0"), val = bool(true)];
+            string input_recurrent_activation_0 = const()[name = string("input_recurrent_activation_0"), val = string("sigmoid")];
+            string input_cell_activation_0 = const()[name = string("input_cell_activation_0"), val = string("tanh")];
+            string input_activation_0 = const()[name = string("input_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7645760)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10922624)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14199488)))];
+            tensor<fp16, [1, 1, 640]> input_cast_fp16_0, tensor<fp16, [1, 640]> input_cast_fp16_1, tensor<fp16, [1, 640]> input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_3_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_lstm_layer_0_cast_fp16_0)[name = string("input_cast_fp16")];
+            int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_cast_fp16_1, input_cast_fp16_1))[name = string("obj_3_cast_fp16")];
+            string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_lstm_layer_0_cast_fp16_2, input_cast_fp16_2))[name = string("obj_cast_fp16")];
+            string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
+            string transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = string("transpose_0_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp16, [1, 640, 1]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = string("transpose_1")];
+            tensor<fp32, [1, 640, 1]> decoder_out = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = string("cast_3")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")];
+        } -> (decoder_out, h_out, c_out);
+}
\ No newline at end of file
diff --git a/fr/2240ms/decoder.mlmodelc/weights/weight.bin b/fr/2240ms/decoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..82ea1696ddad052267574c52c3e3b0bccde479f5
--- /dev/null
+++ b/fr/2240ms/decoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ef2aeef85ed664eb60850330a014c5c299325ba1047d690654dc7ffa870e95a
+size 14204672
diff --git a/fr/2240ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/2240ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..c09a4248bce6ae30d0023f4ec349ae462e19aad7
--- /dev/null
+++ b/fr/2240ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c82c0fbdef3af45d82f9b1fe84c0c822154020a5b3cd0e5070d33e9ef0beeca6
+size 10359
diff --git a/fr/2240ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/2240ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..82ea1696ddad052267574c52c3e3b0bccde479f5
--- /dev/null
+++ b/fr/2240ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ef2aeef85ed664eb60850330a014c5c299325ba1047d690654dc7ffa870e95a
+size 14204672
diff --git a/fr/2240ms/decoder.mlpackage/Manifest.json b/fr/2240ms/decoder.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..80bff9abaf12e957f68dbde257eaf82079fed11b
--- /dev/null
+++ b/fr/2240ms/decoder.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "08617906-EB26-428B-BCF5-BC6013603E2C": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "89EF9907-E7AF-4633-B6D7-C8D472B636BC": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "89EF9907-E7AF-4633-B6D7-C8D472B636BC"
+}
diff --git a/fr/2240ms/decoder_joint.mlmodelc/analytics/coremldata.bin b/fr/2240ms/decoder_joint.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..22debc561a233f97bd7fde99dc9d1a31fb0231de
--- /dev/null
+++ b/fr/2240ms/decoder_joint.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f8367de6fb06c7b8053f5ecbbc8e4c7e5100f85f1e1c9ba6e9d9b3e31ecca3d
+size 243
diff --git a/fr/2240ms/decoder_joint.mlmodelc/coremldata.bin b/fr/2240ms/decoder_joint.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..54e005ff86e2274927ae7ef36a7e0ee4bb128c3b
--- /dev/null
+++ b/fr/2240ms/decoder_joint.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0566656fc89cd81234bb22159c3ccebfb8c0277e3b740d12a8fb53288c1efcd
+size 454
diff --git a/fr/2240ms/decoder_joint.mlmodelc/model.mil b/fr/2240ms/decoder_joint.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..1438ac2ef0ec1f9fc664a6752938e369f7de8e9d
--- /dev/null
+++ b/fr/2240ms/decoder_joint.mlmodelc/model.mil
@@ -0,0 +1,83 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [1, 1024, 1]> encoder, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            int32 y_axis_0 = const()[name = string("y_axis_0"), val = int32(0)];
+            int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)];
+            bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [849, 640]> decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_9")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)];
+            int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)];
+            string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_8")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")];
+            int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)];
+            int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)];
+            string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")];
+            bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)];
+            string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")];
+            string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1086848)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4363712)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7640576)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = string("transpose_4")];
+            tensor<fp16, [1, 1, 640]> input_5_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")];
+            string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")];
+            bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)];
+            string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")];
+            string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7645760)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10922624)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14199488)))];
+            tensor<fp16, [1, 1, 640]> input_5_cast_fp16_0, tensor<fp16, [1, 640]> input_5_cast_fp16_1, tensor<fp16, [1, 640]> input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")];
+            int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")];
+            string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")];
+            string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [3]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [3]> input_7_perm_0 = const()[name = string("input_7_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [640, 1024]> joint_module_enc_weight_to_fp16 = const()[name = string("joint_module_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14204672)))];
+            tensor<fp16, [640]> joint_module_enc_bias_to_fp16 = const()[name = string("joint_module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15515456)))];
+            tensor<fp16, [1, 1024, 1]> encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_4")];
+            tensor<fp16, [1, 1, 1024]> input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = encoder_to_fp16)[name = string("transpose_2")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15516800)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16336064)))];
+            tensor<fp16, [1, 1, 640]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_3")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_79_axes_0 = const()[name = string("op_79_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_79_cast_fp16 = expand_dims(axes = var_79_axes_0, x = linear_0_cast_fp16)[name = string("op_79_cast_fp16")];
+            tensor<int32, [1]> var_80_axes_0 = const()[name = string("op_80_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_80_cast_fp16 = expand_dims(axes = var_80_axes_0, x = linear_1_cast_fp16)[name = string("op_80_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_11_cast_fp16 = add(x = var_79_cast_fp16, y = var_80_cast_fp16)[name = string("input_11_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_13_cast_fp16 = relu(x = input_11_cast_fp16)[name = string("input_13_cast_fp16")];
+            tensor<fp16, [849, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16337408)))];
+            tensor<fp16, [849]> joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [849]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17424192)))];
+            tensor<fp16, [1, 1, 1, 849]> linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_2_cast_fp16")];
+            string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 1, 1, 849]> logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_3")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_5")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_6")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")];
+        } -> (logits, h_out, c_out);
+}
\ No newline at end of file
diff --git a/fr/2240ms/decoder_joint.mlmodelc/weights/weight.bin b/fr/2240ms/decoder_joint.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..7e387cabbe194b0f606c14bdf2b03b19d697a8f3
--- /dev/null
+++ b/fr/2240ms/decoder_joint.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39ce7eecdfb6a432f391c931412d9a5cd15fb0ef85562ef0c207e4a18fe6b50e
+size 17425954
diff --git a/fr/2240ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/2240ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..8de2f672a8b40625e6b0f13654d4cba788880b07
--- /dev/null
+++ b/fr/2240ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72ce6705a73b8b0488f31dd35a385908473aff8bef0ffe290f56bb51068a1d26
+size 13745
diff --git a/fr/2240ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/2240ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..7e387cabbe194b0f606c14bdf2b03b19d697a8f3
--- /dev/null
+++ b/fr/2240ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39ce7eecdfb6a432f391c931412d9a5cd15fb0ef85562ef0c207e4a18fe6b50e
+size 17425954
diff --git a/fr/2240ms/decoder_joint.mlpackage/Manifest.json b/fr/2240ms/decoder_joint.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..967aa622f9a5ac2f95978e9feda32d61c8a73643
--- /dev/null
+++ b/fr/2240ms/decoder_joint.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "8FCEBC80-AF4D-42E1-B70D-96BB47913BC7": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "A77788EC-CA36-4DF3-BB1B-E6182E0ABF28": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "8FCEBC80-AF4D-42E1-B70D-96BB47913BC7"
+}
diff --git a/fr/2240ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin b/fr/2240ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..425b9fbc36cf7602d11ebba8a2e29be972229238
--- /dev/null
+++ b/fr/2240ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:878825967d0c03e7b016222a7b824badb35d1d7e134ad3a973abc00dc5fe353a
+size 243
diff --git a/fr/2240ms/decoder_joint_noencproj.mlmodelc/coremldata.bin b/fr/2240ms/decoder_joint_noencproj.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..479d8bc5a3eab6af7208785446fe73a569a877dc
--- /dev/null
+++ b/fr/2240ms/decoder_joint_noencproj.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ebb7868572e2691055cf8d926fe61e2049bdc0a3e6349e82227f0b82dd849814
+size 519
diff --git a/fr/2240ms/decoder_joint_noencproj.mlmodelc/model.mil b/fr/2240ms/decoder_joint_noencproj.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..ef9fa8f205f75bf414b61d0b622d2976b441fe5b
--- /dev/null
+++ b/fr/2240ms/decoder_joint_noencproj.mlmodelc/model.mil
@@ -0,0 +1,91 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios18>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [1, 1, 640]> encoder_proj, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)];
+            bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [849, 640]> decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")];
+            string cast_1_dtype_0 = const()[name = string("cast_1_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_10")];
+            tensor<int32, [1, 1]> cast_1 = cast(dtype = cast_1_dtype_0, x = token_to_int16)[name = string("cast_9")];
+            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
+            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(849)];
+            tensor<int32, [1, 1]> add_2 = add(x = cast_1, y = slice_by_index_0)[name = string("add_2")];
+            tensor<int32, [1, 1]> select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = string("select_0")];
+            int32 y_cast_fp16_cast_uint16_axis_0 = const()[name = string("y_cast_fp16_cast_uint16_axis_0"), val = int32(0)];
+            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
+            tensor<int16, [1, 1]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_8")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)];
+            int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)];
+            string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")];
+            int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)];
+            int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)];
+            string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")];
+            bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)];
+            string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")];
+            string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1086848)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4363712)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7640576)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = string("transpose_3")];
+            tensor<fp16, [1, 1, 640]> input_5_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")];
+            string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")];
+            bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)];
+            string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")];
+            string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7645760)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10922624)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14199488)))];
+            tensor<fp16, [1, 1, 640]> input_5_cast_fp16_0, tensor<fp16, [1, 640]> input_5_cast_fp16_1, tensor<fp16, [1, 640]> input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")];
+            int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")];
+            string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")];
+            string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [3]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14204672)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15023936)))];
+            tensor<fp16, [1, 1, 640]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_2")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [1]> f_axes_0 = const()[name = string("f_axes_0"), val = tensor<int32, [1]>([2])];
+            string encoder_proj_to_fp16_dtype_0 = const()[name = string("encoder_proj_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 1, 640]> encoder_proj_to_fp16 = cast(dtype = encoder_proj_to_fp16_dtype_0, x = encoder_proj)[name = string("cast_3")];
+            tensor<fp16, [1, 1, 1, 640]> f_cast_fp16 = expand_dims(axes = f_axes_0, x = encoder_proj_to_fp16)[name = string("f_cast_fp16")];
+            tensor<int32, [1]> g_axes_0 = const()[name = string("g_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> g_cast_fp16 = expand_dims(axes = g_axes_0, x = linear_0_cast_fp16)[name = string("g_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_9_cast_fp16 = add(x = f_cast_fp16, y = g_cast_fp16)[name = string("input_9_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_11_cast_fp16 = relu(x = input_9_cast_fp16)[name = string("input_11_cast_fp16")];
+            tensor<fp16, [849, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15025280)))];
+            tensor<fp16, [849]> joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [849]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16112064)))];
+            tensor<fp16, [1, 1, 1, 849]> linear_1_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_1_cast_fp16")];
+            int32 var_83 = const()[name = string("op_83"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1, 849]> var_85_softmax_cast_fp16 = softmax(axis = var_83, x = linear_1_cast_fp16)[name = string("op_85_softmax_cast_fp16")];
+            fp32 var_85_epsilon_0 = const()[name = string("op_85_epsilon_0"), val = fp32(0x1p-149)];
+            tensor<fp16, [1, 1, 1, 849]> var_85_cast_fp16 = log(epsilon = var_85_epsilon_0, x = var_85_softmax_cast_fp16)[name = string("op_85_cast_fp16")];
+            string var_85_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_85_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 1, 1, 849]> logits = cast(dtype = var_85_cast_fp16_to_fp32_dtype_0, x = var_85_cast_fp16)[name = string("cast_2")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")];
+        } -> (logits, h_out, c_out);
+}
\ No newline at end of file
diff --git a/fr/2240ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin b/fr/2240ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..e37358fc0d446e079fe87526671dc411c4b37591
--- /dev/null
+++ b/fr/2240ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c691d3004f2ac6a956208569a29d899a7f8e5030087d53fbd488785fa0e8efad
+size 16113826
diff --git a/fr/2240ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/2240ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..b701505b2faf5efd4a3c1024077f4834b1f94d37
--- /dev/null
+++ b/fr/2240ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4196a821e716db75328fdd2f99f83214ecd9e6e6481a96a88907fe1c4232f143
+size 14630
diff --git a/fr/2240ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/2240ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..e37358fc0d446e079fe87526671dc411c4b37591
--- /dev/null
+++ b/fr/2240ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c691d3004f2ac6a956208569a29d899a7f8e5030087d53fbd488785fa0e8efad
+size 16113826
diff --git a/fr/2240ms/decoder_joint_noencproj.mlpackage/Manifest.json b/fr/2240ms/decoder_joint_noencproj.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..4709121d3572643a899871aa076a46f4c02a7ef8
--- /dev/null
+++ b/fr/2240ms/decoder_joint_noencproj.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "1D4938C1-F035-4D52-9376-C0FB72335508": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "BF28BA0A-268B-4C46-AB97-C0173D8B55F0": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "BF28BA0A-268B-4C46-AB97-C0173D8B55F0"
+}
diff --git a/fr/2240ms/encoder.mlmodelc/analytics/coremldata.bin b/fr/2240ms/encoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..cf222b05b3795ceca6d95b114574c50f82d5a9fc
--- /dev/null
+++ b/fr/2240ms/encoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cb008a0254143d32e93073d5db48272287719969c062c9ac55514b11dc699a3f
+size 243
diff --git a/fr/2240ms/encoder.mlmodelc/coremldata.bin b/fr/2240ms/encoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..1f6cce0682fde74d9acea28989562b41e7722121
--- /dev/null
+++ b/fr/2240ms/encoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5fa8f4f96d6e67d4b14d239fb22e61f909f789d9dd15501f084ce27a67a6aff3
+size 633
diff --git a/fr/2240ms/encoder.mlmodelc/model.mil b/fr/2240ms/encoder.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..0b61772b118bff8a3d2b83d31f02ad558e04cde4
--- /dev/null
+++ b/fr/2240ms/encoder.mlmodelc/model.mil
@@ -0,0 +1,4434 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})]
+{
+    func main<ios18>(tensor<fp32, [1, 24, 42, 1024]> cache_channel, tensor<int32, [1]> cache_len, tensor<fp32, [1, 24, 1024, 8]> cache_time, tensor<fp32, [1, 128, 233]> mel, tensor<int32, [1]> mel_length, tensor<int32, [1]> prompt_id) {
+            tensor<int32, [4]> value_3_perm_0 = const()[name = string("value_3_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string cache_channel_to_fp16_dtype_0 = const()[name = string("cache_channel_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<int32, [4]> value_5_perm_0 = const()[name = string("value_5_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string cache_time_to_fp16_dtype_0 = const()[name = string("cache_time_to_fp16_dtype_0"), val = string("fp16")];
+            int32 var_59 = const()[name = string("op_59"), val = int32(-1)];
+            int32 var_68 = const()[name = string("op_68"), val = int32(1)];
+            tensor<int32, [3]> x_1_perm_0 = const()[name = string("x_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string mel_to_fp16_dtype_0 = const()[name = string("mel_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<int32, [1]> tensor_1_axes_0 = const()[name = string("tensor_1_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 128, 233]> mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = string("cast_21")];
+            tensor<fp16, [1, 233, 128]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = mel_to_fp16)[name = string("transpose_367")];
+            tensor<fp16, [1, 1, 233, 128]> tensor_1_cast_fp16 = expand_dims(axes = tensor_1_axes_0, x = x_1_cast_fp16)[name = string("tensor_1_cast_fp16")];
+            tensor<int32, [1, 233]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1, 233]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<int32, [1]> var_137_axes_0 = const()[name = string("op_137_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_137 = expand_dims(axes = var_137_axes_0, x = mel_length)[name = string("op_137")];
+            tensor<bool, [1, 233]> time_mask_1 = less(x = expand_dims_0, y = var_137)[name = string("time_mask_1")];
+            tensor<int32, [1]> var_139_axes_0 = const()[name = string("op_139_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 233, 1]> var_139 = expand_dims(axes = var_139_axes_0, x = time_mask_1)[name = string("op_139")];
+            tensor<int32, [3]> var_141_reps_0 = const()[name = string("op_141_reps_0"), val = tensor<int32, [3]>([1, 1, 128])];
+            tensor<bool, [1, 233, 128]> var_141 = tile(reps = var_141_reps_0, x = var_139)[name = string("op_141")];
+            tensor<int32, [1]> var_147_axes_0 = const()[name = string("op_147_axes_0"), val = tensor<int32, [1]>([1])];
+            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 233, 128]> var_141_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_141)[name = string("cast_20")];
+            tensor<fp16, [1, 1, 233, 128]> var_147_cast_fp16 = expand_dims(axes = var_147_axes_0, x = var_141_to_fp16)[name = string("op_147_cast_fp16")];
+            tensor<fp16, [1, 1, 233, 128]> input_1_cast_fp16 = mul(x = tensor_1_cast_fp16, y = var_147_cast_fp16)[name = string("input_1_cast_fp16")];
+            tensor<int32, [8]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("constant")];
+            fp16 const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 236, 131]> input_3_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")];
+            string tensor_3_pad_type_0 = const()[name = string("tensor_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_3_strides_0 = const()[name = string("tensor_3_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, [4]> tensor_3_pad_0 = const()[name = string("tensor_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_3_dilations_0 = const()[name = string("tensor_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 tensor_3_groups_0 = const()[name = string("tensor_3_groups_0"), val = int32(1)];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_0_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1088))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3456))))[name = string("encoder_pre_encode_conv_0_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4032)))];
+            tensor<fp16, [1, 256, 117, 65]> tensor_3_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = tensor_3_dilations_0, groups = tensor_3_groups_0, pad = tensor_3_pad_0, pad_type = tensor_3_pad_type_0, strides = tensor_3_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("tensor_3_cast_fp16")];
+            string current_lengths_1_to_fp16_dtype_0 = const()[name = string("current_lengths_1_to_fp16_dtype_0"), val = string("fp16")];
+            fp16 var_160_promoted_to_fp16 = const()[name = string("op_160_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> mel_length_to_fp16 = cast(dtype = current_lengths_1_to_fp16_dtype_0, x = mel_length)[name = string("cast_19")];
+            tensor<fp16, [1]> var_161_cast_fp16 = add(x = mel_length_to_fp16, y = var_160_promoted_to_fp16)[name = string("op_161_cast_fp16")];
+            fp16 var_162_promoted_to_fp16 = const()[name = string("op_162_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> var_163_cast_fp16 = add(x = var_161_cast_fp16, y = var_162_promoted_to_fp16)[name = string("op_163_cast_fp16")];
+            fp16 var_164_promoted_to_fp16 = const()[name = string("op_164_promoted_to_fp16"), val = fp16(0x1.8p+1)];
+            tensor<fp16, [1]> var_165_cast_fp16 = sub(x = var_163_cast_fp16, y = var_164_promoted_to_fp16)[name = string("op_165_cast_fp16")];
+            fp16 var_56_promoted_to_fp16 = const()[name = string("op_56_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> floor_div_0_cast_fp16 = floor_div(x = var_165_cast_fp16, y = var_56_promoted_to_fp16)[name = string("floor_div_0_cast_fp16")];
+            fp16 var_167_promoted_to_fp16 = const()[name = string("op_167_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_3_cast_fp16 = add(x = floor_div_0_cast_fp16, y = var_167_promoted_to_fp16)[name = string("current_lengths_3_cast_fp16")];
+            string lengths_19_dtype_0 = const()[name = string("lengths_19_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 117]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1, 117]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4608)))];
+            tensor<int32, [1]> var_176_axes_0 = const()[name = string("op_176_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_3_cast_fp16_to_int32 = cast(dtype = lengths_19_dtype_0, x = current_lengths_3_cast_fp16)[name = string("cast_18")];
+            tensor<int32, [1, 1]> var_176 = expand_dims(axes = var_176_axes_0, x = current_lengths_3_cast_fp16_to_int32)[name = string("op_176")];
+            tensor<bool, [1, 117]> time_mask_3 = less(x = expand_dims_1, y = var_176)[name = string("time_mask_3")];
+            tensor<int32, [1]> var_178_axes_0 = const()[name = string("op_178_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 117, 1]> var_178 = expand_dims(axes = var_178_axes_0, x = time_mask_3)[name = string("op_178")];
+            tensor<int32, [3]> var_180_reps_0 = const()[name = string("op_180_reps_0"), val = tensor<int32, [3]>([1, 1, 65])];
+            tensor<bool, [1, 117, 65]> var_180 = tile(reps = var_180_reps_0, x = var_178)[name = string("op_180")];
+            tensor<int32, [1]> var_186_axes_0 = const()[name = string("op_186_axes_0"), val = tensor<int32, [1]>([1])];
+            string mask_3_to_fp16_dtype_0 = const()[name = string("mask_3_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 117, 65]> var_180_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_180)[name = string("cast_17")];
+            tensor<fp16, [1, 1, 117, 65]> var_186_cast_fp16 = expand_dims(axes = var_186_axes_0, x = var_180_to_fp16)[name = string("op_186_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_3_reps_0 = const()[name = string("expanded_mask_3_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 117, 65]> expanded_mask_3_cast_fp16 = tile(reps = expanded_mask_3_reps_0, x = var_186_cast_fp16)[name = string("expanded_mask_3_cast_fp16")];
+            tensor<fp16, [1, 256, 117, 65]> input_5_cast_fp16 = mul(x = tensor_3_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<fp16, [1, 256, 117, 65]> tensor_5_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("tensor_5_cast_fp16")];
+            tensor<fp16, [1, 256, 117, 65]> input_7_cast_fp16 = mul(x = tensor_5_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_7_cast_fp16")];
+            tensor<int32, [8]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            string input_9_mode_0 = const()[name = string("input_9_mode_0"), val = string("constant")];
+            fp16 const_23_to_fp16 = const()[name = string("const_23_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 256, 120, 68]> input_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")];
+            string tensor_7_pad_type_0 = const()[name = string("tensor_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_7_strides_0 = const()[name = string("tensor_7_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            int32 tensor_7_groups_0 = const()[name = string("tensor_7_groups_0"), val = int32(256)];
+            tensor<int32, [4]> tensor_7_pad_0 = const()[name = string("tensor_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_7_dilations_0 = const()[name = string("tensor_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5184))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7552))))[name = string("encoder_pre_encode_conv_2_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8128)))];
+            tensor<fp16, [1, 256, 59, 33]> tensor_7_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = tensor_7_dilations_0, groups = tensor_7_groups_0, pad = tensor_7_pad_0, pad_type = tensor_7_pad_type_0, strides = tensor_7_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("tensor_7_cast_fp16")];
+            fp16 var_208_promoted_to_fp16 = const()[name = string("op_208_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> var_209_cast_fp16 = add(x = current_lengths_3_cast_fp16, y = var_208_promoted_to_fp16)[name = string("op_209_cast_fp16")];
+            fp16 var_210_promoted_to_fp16 = const()[name = string("op_210_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> var_211_cast_fp16 = add(x = var_209_cast_fp16, y = var_210_promoted_to_fp16)[name = string("op_211_cast_fp16")];
+            fp16 var_212_promoted_to_fp16 = const()[name = string("op_212_promoted_to_fp16"), val = fp16(0x1.8p+1)];
+            tensor<fp16, [1]> var_213_cast_fp16 = sub(x = var_211_cast_fp16, y = var_212_promoted_to_fp16)[name = string("op_213_cast_fp16")];
+            fp16 var_56_promoted_1_to_fp16 = const()[name = string("op_56_promoted_1_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> floor_div_1_cast_fp16 = floor_div(x = var_213_cast_fp16, y = var_56_promoted_1_to_fp16)[name = string("floor_div_1_cast_fp16")];
+            fp16 var_215_promoted_to_fp16 = const()[name = string("op_215_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_5_cast_fp16 = add(x = floor_div_1_cast_fp16, y = var_215_promoted_to_fp16)[name = string("current_lengths_5_cast_fp16")];
+            string lengths_21_dtype_0 = const()[name = string("lengths_21_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 59]> expand_dims_2 = const()[name = string("expand_dims_2"), val = tensor<int32, [1, 59]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8704)))];
+            tensor<int32, [1]> var_224_axes_0 = const()[name = string("op_224_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_5_cast_fp16_to_int32 = cast(dtype = lengths_21_dtype_0, x = current_lengths_5_cast_fp16)[name = string("cast_16")];
+            tensor<int32, [1, 1]> var_224 = expand_dims(axes = var_224_axes_0, x = current_lengths_5_cast_fp16_to_int32)[name = string("op_224")];
+            tensor<bool, [1, 59]> time_mask_5 = less(x = expand_dims_2, y = var_224)[name = string("time_mask_5")];
+            tensor<int32, [1]> var_226_axes_0 = const()[name = string("op_226_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 59, 1]> var_226 = expand_dims(axes = var_226_axes_0, x = time_mask_5)[name = string("op_226")];
+            tensor<int32, [3]> var_228_reps_0 = const()[name = string("op_228_reps_0"), val = tensor<int32, [3]>([1, 1, 33])];
+            tensor<bool, [1, 59, 33]> var_228 = tile(reps = var_228_reps_0, x = var_226)[name = string("op_228")];
+            tensor<int32, [1]> var_234_axes_0 = const()[name = string("op_234_axes_0"), val = tensor<int32, [1]>([1])];
+            string mask_5_to_fp16_dtype_0 = const()[name = string("mask_5_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 59, 33]> var_228_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_228)[name = string("cast_15")];
+            tensor<fp16, [1, 1, 59, 33]> var_234_cast_fp16 = expand_dims(axes = var_234_axes_0, x = var_228_to_fp16)[name = string("op_234_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_7_reps_0 = const()[name = string("expanded_mask_7_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 59, 33]> expanded_mask_7_cast_fp16 = tile(reps = expanded_mask_7_reps_0, x = var_234_cast_fp16)[name = string("expanded_mask_7_cast_fp16")];
+            tensor<fp16, [1, 256, 59, 33]> input_11_cast_fp16 = mul(x = tensor_7_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_11_cast_fp16")];
+            string tensor_9_pad_type_0 = const()[name = string("tensor_9_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_9_strides_0 = const()[name = string("tensor_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> tensor_9_pad_0 = const()[name = string("tensor_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_9_dilations_0 = const()[name = string("tensor_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 tensor_9_groups_0 = const()[name = string("tensor_9_groups_0"), val = int32(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_3_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9024))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74624))))[name = string("encoder_pre_encode_conv_3_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75200)))];
+            tensor<fp16, [1, 256, 59, 33]> tensor_9_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = tensor_9_dilations_0, groups = tensor_9_groups_0, pad = tensor_9_pad_0, pad_type = tensor_9_pad_type_0, strides = tensor_9_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("tensor_9_cast_fp16")];
+            tensor<fp16, [1, 256, 59, 33]> input_13_cast_fp16 = mul(x = tensor_9_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_13_cast_fp16")];
+            tensor<fp16, [1, 256, 59, 33]> tensor_11_cast_fp16 = relu(x = input_13_cast_fp16)[name = string("tensor_11_cast_fp16")];
+            tensor<fp16, [1, 256, 59, 33]> input_15_cast_fp16 = mul(x = tensor_11_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_15_cast_fp16")];
+            tensor<int32, [8]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            string input_17_mode_0 = const()[name = string("input_17_mode_0"), val = string("constant")];
+            fp16 const_41_to_fp16 = const()[name = string("const_41_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 256, 62, 36]> input_17_cast_fp16 = pad(constant_val = const_41_to_fp16, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")];
+            string tensor_13_pad_type_0 = const()[name = string("tensor_13_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_13_strides_0 = const()[name = string("tensor_13_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            int32 tensor_13_groups_0 = const()[name = string("tensor_13_groups_0"), val = int32(256)];
+            tensor<int32, [4]> tensor_13_pad_0 = const()[name = string("tensor_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_13_dilations_0 = const()[name = string("tensor_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_5_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75776))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78144))))[name = string("encoder_pre_encode_conv_5_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78720)))];
+            tensor<fp16, [1, 256, 30, 17]> tensor_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = tensor_13_dilations_0, groups = tensor_13_groups_0, pad = tensor_13_pad_0, pad_type = tensor_13_pad_type_0, strides = tensor_13_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("tensor_13_cast_fp16")];
+            fp16 var_271_promoted_to_fp16 = const()[name = string("op_271_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> var_272_cast_fp16 = add(x = current_lengths_5_cast_fp16, y = var_271_promoted_to_fp16)[name = string("op_272_cast_fp16")];
+            fp16 var_273_promoted_to_fp16 = const()[name = string("op_273_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> var_274_cast_fp16 = add(x = var_272_cast_fp16, y = var_273_promoted_to_fp16)[name = string("op_274_cast_fp16")];
+            fp16 var_275_promoted_to_fp16 = const()[name = string("op_275_promoted_to_fp16"), val = fp16(0x1.8p+1)];
+            tensor<fp16, [1]> var_276_cast_fp16 = sub(x = var_274_cast_fp16, y = var_275_promoted_to_fp16)[name = string("op_276_cast_fp16")];
+            fp16 var_56_promoted_2_to_fp16 = const()[name = string("op_56_promoted_2_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> floor_div_2_cast_fp16 = floor_div(x = var_276_cast_fp16, y = var_56_promoted_2_to_fp16)[name = string("floor_div_2_cast_fp16")];
+            fp16 var_278_promoted_to_fp16 = const()[name = string("op_278_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_cast_fp16 = add(x = floor_div_2_cast_fp16, y = var_278_promoted_to_fp16)[name = string("current_lengths_cast_fp16")];
+            string lengths_dtype_0 = const()[name = string("lengths_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 30]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1, 30]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79296)))];
+            tensor<int32, [1]> var_287_axes_0 = const()[name = string("op_287_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_cast_fp16_to_int32 = cast(dtype = lengths_dtype_0, x = current_lengths_cast_fp16)[name = string("cast_14")];
+            tensor<int32, [1, 1]> var_287 = expand_dims(axes = var_287_axes_0, x = current_lengths_cast_fp16_to_int32)[name = string("op_287")];
+            tensor<bool, [1, 30]> time_mask = less(x = expand_dims_3, y = var_287)[name = string("time_mask")];
+            tensor<int32, [1]> var_289_axes_0 = const()[name = string("op_289_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 30, 1]> var_289 = expand_dims(axes = var_289_axes_0, x = time_mask)[name = string("op_289")];
+            tensor<int32, [3]> var_291_reps_0 = const()[name = string("op_291_reps_0"), val = tensor<int32, [3]>([1, 1, 17])];
+            tensor<bool, [1, 30, 17]> var_291 = tile(reps = var_291_reps_0, x = var_289)[name = string("op_291")];
+            tensor<int32, [1]> var_297_axes_0 = const()[name = string("op_297_axes_0"), val = tensor<int32, [1]>([1])];
+            string mask_7_to_fp16_dtype_0 = const()[name = string("mask_7_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 30, 17]> var_291_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_291)[name = string("cast_13")];
+            tensor<fp16, [1, 1, 30, 17]> var_297_cast_fp16 = expand_dims(axes = var_297_axes_0, x = var_291_to_fp16)[name = string("op_297_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_13_reps_0 = const()[name = string("expanded_mask_13_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 30, 17]> expanded_mask_13_cast_fp16 = tile(reps = expanded_mask_13_reps_0, x = var_297_cast_fp16)[name = string("expanded_mask_13_cast_fp16")];
+            tensor<fp16, [1, 256, 30, 17]> input_19_cast_fp16 = mul(x = tensor_13_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_19_cast_fp16")];
+            string tensor_15_pad_type_0 = const()[name = string("tensor_15_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_15_strides_0 = const()[name = string("tensor_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> tensor_15_pad_0 = const()[name = string("tensor_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_15_dilations_0 = const()[name = string("tensor_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 tensor_15_groups_0 = const()[name = string("tensor_15_groups_0"), val = int32(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_6_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79488))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145088))))[name = string("encoder_pre_encode_conv_6_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145664)))];
+            tensor<fp16, [1, 256, 30, 17]> tensor_15_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = tensor_15_dilations_0, groups = tensor_15_groups_0, pad = tensor_15_pad_0, pad_type = tensor_15_pad_type_0, strides = tensor_15_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("tensor_15_cast_fp16")];
+            tensor<fp16, [1, 256, 30, 17]> input_21_cast_fp16 = mul(x = tensor_15_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_21_cast_fp16")];
+            tensor<fp16, [1, 256, 30, 17]> tensor_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("tensor_cast_fp16")];
+            tensor<fp16, [1, 256, 30, 17]> x_3_cast_fp16 = mul(x = tensor_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("x_3_cast_fp16")];
+            tensor<int32, [4]> var_331_perm_0 = const()[name = string("op_331_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_332 = const()[name = string("op_332"), val = tensor<int32, [3]>([1, 30, -1])];
+            tensor<fp16, [1, 30, 256, 17]> var_331_cast_fp16 = transpose(perm = var_331_perm_0, x = x_3_cast_fp16)[name = string("transpose_366")];
+            tensor<fp16, [1, 30, 4352]> input_23_cast_fp16 = reshape(shape = var_332, x = var_331_cast_fp16)[name = string("input_23_cast_fp16")];
+            tensor<fp16, [1024, 4352]> encoder_pre_encode_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4352]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146240))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4602752))))[name = string("encoder_pre_encode_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_pre_encode_out_bias_to_fp16 = const()[name = string("encoder_pre_encode_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4604864)))];
+            tensor<fp16, [1, 30, 1024]> linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [3]> var_342_begin_0 = const()[name = string("op_342_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_342_end_0 = const()[name = string("op_342_end_0"), val = tensor<int32, [3]>([1, 30, 1024])];
+            tensor<bool, [3]> var_342_end_mask_0 = const()[name = string("op_342_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 28, 1024]> var_342_cast_fp16 = slice_by_index(begin = var_342_begin_0, end = var_342_end_0, end_mask = var_342_end_mask_0, x = linear_0_cast_fp16)[name = string("op_342_cast_fp16")];
+            int32 var_344 = const()[name = string("op_344"), val = int32(2)];
+            tensor<int32, [1]> var_345 = sub(x = current_lengths_cast_fp16_to_int32, y = var_344)[name = string("op_345")];
+            string var_345_promoted_to_fp16_dtype_0 = const()[name = string("op_345_promoted_to_fp16_dtype_0"), val = string("fp16")];
+            fp16 var_62_promoted_to_fp16 = const()[name = string("op_62_promoted_to_fp16"), val = fp16(0x0p+0)];
+            fp16 const_61_to_fp16 = const()[name = string("const_61_to_fp16"), val = fp16(inf)];
+            tensor<fp16, [1]> var_345_to_fp16 = cast(dtype = var_345_promoted_to_fp16_dtype_0, x = var_345)[name = string("cast_12")];
+            tensor<fp16, [1]> clip_0_cast_fp16 = clip(alpha = var_62_promoted_to_fp16, beta = const_61_to_fp16, x = var_345_to_fp16)[name = string("clip_0_cast_fp16")];
+            tensor<int32, [1]> max_audio_length_1 = const()[name = string("max_audio_length_1"), val = tensor<int32, [1]>([28])];
+            fp16 var_361_promoted_to_fp16 = const()[name = string("op_361_promoted_to_fp16"), val = fp16(0x1.5p+5)];
+            tensor<fp16, [1]> padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_361_promoted_to_fp16)[name = string("padding_length_cast_fp16")];
+            int32 const_63 = const()[name = string("const_63"), val = int32(-1)];
+            tensor<int32, [1]> var_363 = mul(x = cache_len, y = const_63)[name = string("op_363")];
+            int32 var_364 = const()[name = string("op_364"), val = int32(42)];
+            tensor<int32, [1]> offset = add(x = var_363, y = var_364)[name = string("offset")];
+            tensor<int32, [1]> var_404_axes_0 = const()[name = string("op_404_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 1]> var_404_cast_fp16 = expand_dims(axes = var_404_axes_0, x = padding_length_cast_fp16)[name = string("op_404_cast_fp16")];
+            tensor<fp16, [1, 70]> var_403_promoted_to_fp16 = const()[name = string("op_403_promoted_to_fp16"), val = tensor<fp16, [1, 70]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606976)))];
+            tensor<bool, [1, 70]> pad_mask_1_cast_fp16 = less(x = var_403_promoted_to_fp16, y = var_404_cast_fp16)[name = string("pad_mask_1_cast_fp16")];
+            tensor<int32, [1, 70]> expand_dims_5 = const()[name = string("expand_dims_5"), val = tensor<int32, [1, 70]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4607232)))];
+            tensor<int32, [1]> var_410_axes_0 = const()[name = string("op_410_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 1]> var_410 = expand_dims(axes = var_410_axes_0, x = offset)[name = string("op_410")];
+            tensor<bool, [1, 70]> pad_mask_off = greater_equal(x = expand_dims_5, y = var_410)[name = string("pad_mask_off")];
+            tensor<bool, [1, 70]> pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = string("pad_mask_3")];
+            tensor<int32, [1]> var_413_axes_0 = const()[name = string("op_413_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 70]> var_413 = expand_dims(axes = var_413_axes_0, x = pad_mask_3)[name = string("op_413")];
+            tensor<int32, [3]> var_414 = const()[name = string("op_414"), val = tensor<int32, [3]>([1, 70, 1])];
+            tensor<bool, [1, 70, 70]> pad_mask_for_att_mask_1 = tile(reps = var_414, x = var_413)[name = string("pad_mask_for_att_mask_1")];
+            tensor<int32, [3]> var_416_perm_0 = const()[name = string("op_416_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<bool, [1, 70, 70]> var_416 = transpose(perm = var_416_perm_0, x = pad_mask_for_att_mask_1)[name = string("transpose_365")];
+            tensor<bool, [1, 70, 70]> pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_416)[name = string("pad_mask_for_att_mask")];
+            tensor<bool, [1, 70, 70]> const_71 = const()[name = string("const_71"), val = tensor<bool, [1, 70, 70]>([[[true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])];
+            tensor<bool, [1, 70, 70]> att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_71)[name = string("att_mask_9")];
+            tensor<bool, [1, 70, 70]> att_mask = logical_not(x = att_mask_9)[name = string("att_mask")];
+            tensor<bool, [1, 70]> pad_mask_5 = logical_not(x = pad_mask_3)[name = string("pad_mask_5")];
+            tensor<int32, [2]> pad_mask_begin_0 = const()[name = string("pad_mask_begin_0"), val = tensor<int32, [2]>([0, 42])];
+            tensor<int32, [2]> pad_mask_end_0 = const()[name = string("pad_mask_end_0"), val = tensor<int32, [2]>([1, 70])];
+            tensor<bool, [2]> pad_mask_end_mask_0 = const()[name = string("pad_mask_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<bool, [1, 28]> pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = string("pad_mask")];
+            tensor<int32, [3]> mask_9_begin_0 = const()[name = string("mask_9_begin_0"), val = tensor<int32, [3]>([0, 42, 0])];
+            tensor<int32, [3]> mask_9_end_0 = const()[name = string("mask_9_end_0"), val = tensor<int32, [3]>([1, 70, 70])];
+            tensor<bool, [3]> mask_9_end_mask_0 = const()[name = string("mask_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<bool, [1, 28, 70]> mask_9 = slice_by_index(begin = mask_9_begin_0, end = mask_9_end_0, end_mask = mask_9_end_mask_0, x = att_mask)[name = string("mask_9")];
+            tensor<int32, [4]> cache_1_begin_0 = const()[name = string("cache_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_1_end_0 = const()[name = string("cache_1_end_0"), val = tensor<int32, [4]>([1, 1, 42, 1024])];
+            tensor<bool, [4]> cache_1_end_mask_0 = const()[name = string("cache_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_1_squeeze_mask_0 = const()[name = string("cache_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 24, 42, 1024]> cache_channel_to_fp16 = cast(dtype = cache_channel_to_fp16_dtype_0, x = cache_channel)[name = string("cast_11")];
+            tensor<fp16, [24, 1, 42, 1024]> value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = cache_channel_to_fp16)[name = string("transpose_364")];
+            tensor<fp16, [1, 42, 1024]> cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_1_cast_fp16")];
+            tensor<int32, [4]> cache_3_begin_0 = const()[name = string("cache_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_3_end_0 = const()[name = string("cache_3_end_0"), val = tensor<int32, [4]>([1, 1, 1024, 8])];
+            tensor<bool, [4]> cache_3_end_mask_0 = const()[name = string("cache_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_3_squeeze_mask_0 = const()[name = string("cache_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 24, 1024, 8]> cache_time_to_fp16 = cast(dtype = cache_time_to_fp16_dtype_0, x = cache_time)[name = string("cast_10")];
+            tensor<fp16, [24, 1, 1024, 8]> value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = cache_time_to_fp16)[name = string("transpose_363")];
+            tensor<fp16, [1, 1024, 8]> cache_3_cast_fp16 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_3_cast_fp16")];
+            tensor<int32, [1]> input_27_axes_0 = const()[name = string("input_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4607616)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4609728)))];
+            fp16 var_42_to_fp16 = const()[name = string("op_42_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 28, 1024]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_342_cast_fp16)[name = string("input_27_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4611840))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8806208))))[name = string("encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_0_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8814464)))];
+            tensor<fp16, [1, 28, 4096]> linear_1_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = string("input_31_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8822720))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13017088))))[name = string("encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13019200)))];
+            tensor<fp16, [1, 28, 1024]> linear_2_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_2_cast_fp16")];
+            fp16 var_455_to_fp16 = const()[name = string("op_455_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_456_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_455_to_fp16)[name = string("op_456_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_37_cast_fp16 = add(x = var_342_cast_fp16, y = var_456_cast_fp16)[name = string("input_37_cast_fp16")];
+            tensor<int32, [1]> key_1_axes_0 = const()[name = string("key_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13021312)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13023424)))];
+            tensor<fp16, [1, 28, 1024]> key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = string("key_1_cast_fp16")];
+            bool input_39_interleave_0 = const()[name = string("input_39_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_39_cast_fp16 = concat(axis = var_68, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = string("input_39_cast_fp16")];
+            tensor<int32, [3]> var_478_begin_0 = const()[name = string("op_478_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_478_end_0 = const()[name = string("op_478_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_478_end_mask_0 = const()[name = string("op_478_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_478_cast_fp16 = slice_by_index(begin = var_478_begin_0, end = var_478_end_0, end_mask = var_478_end_mask_0, x = cache_1_cast_fp16)[name = string("op_478_cast_fp16")];
+            bool var_484_interleave_0 = const()[name = string("op_484_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_484_cast_fp16 = concat(axis = var_68, interleave = var_484_interleave_0, values = (var_478_cast_fp16, key_1_cast_fp16))[name = string("op_484_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13025536))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14074176))))[name = string("encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14076288)))];
+            tensor<fp16, [1, 28, 1024]> linear_3_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized, x = key_1_cast_fp16)[name = string("linear_3_cast_fp16")];
+            tensor<int32, [4]> var_489 = const()[name = string("op_489"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_1_cast_fp16 = reshape(shape = var_489, x = linear_3_cast_fp16)[name = string("q_1_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14078400))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15127040))))[name = string("encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15129152)))];
+            tensor<fp16, [1, 70, 1024]> linear_4_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_4_cast_fp16")];
+            tensor<int32, [4]> var_494 = const()[name = string("op_494"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_1_cast_fp16 = reshape(shape = var_494, x = linear_4_cast_fp16)[name = string("k_1_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15131264))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16179904))))[name = string("encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16182016)))];
+            tensor<fp16, [1, 70, 1024]> linear_5_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_5_cast_fp16")];
+            tensor<int32, [4]> var_499 = const()[name = string("op_499"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_1_cast_fp16 = reshape(shape = var_499, x = linear_5_cast_fp16)[name = string("v_1_cast_fp16")];
+            tensor<int32, [4]> value_9_perm_0 = const()[name = string("value_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16184128)))];
+            tensor<fp16, [1, 28, 8, 128]> var_512_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = string("op_512_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16186240)))];
+            tensor<fp16, [1, 28, 8, 128]> var_514_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = string("op_514_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_1_perm_0 = const()[name = string("q_with_bias_v_1_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_7_transpose_x_0 = const()[name = string("x_7_transpose_x_0"), val = bool(false)];
+            bool x_7_transpose_y_0 = const()[name = string("x_7_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_516_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16188352))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16330752))))[name = string("op_516_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_514_cast_fp16)[name = string("transpose_362")];
+            tensor<fp16, [1, 8, 28, 139]> x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = op_516_to_fp16_quantized)[name = string("x_7_cast_fp16")];
+            tensor<int32, [8]> x_9_pad_0 = const()[name = string("x_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_9_mode_0 = const()[name = string("x_9_mode_0"), val = string("constant")];
+            fp16 const_79_to_fp16 = const()[name = string("const_79_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_9_cast_fp16 = pad(constant_val = const_79_to_fp16, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7_cast_fp16)[name = string("x_9_cast_fp16")];
+            tensor<int32, [4]> var_524 = const()[name = string("op_524"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_11_cast_fp16 = reshape(shape = var_524, x = x_9_cast_fp16)[name = string("x_11_cast_fp16")];
+            tensor<int32, [4]> var_528_begin_0 = const()[name = string("op_528_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_528_end_0 = const()[name = string("op_528_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_528_end_mask_0 = const()[name = string("op_528_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_528_cast_fp16 = slice_by_index(begin = var_528_begin_0, end = var_528_end_0, end_mask = var_528_end_mask_0, x = x_11_cast_fp16)[name = string("op_528_cast_fp16")];
+            tensor<int32, [4]> var_529 = const()[name = string("op_529"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_1_cast_fp16 = reshape(shape = var_529, x = var_528_cast_fp16)[name = string("matrix_bd_1_cast_fp16")];
+            bool matrix_ac_1_transpose_x_0 = const()[name = string("matrix_ac_1_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_1_transpose_y_0 = const()[name = string("matrix_ac_1_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_97 = transpose(perm = transpose_97_perm_0, x = k_1_cast_fp16)[name = string("transpose_360")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_96 = transpose(perm = transpose_96_perm_0, x = var_512_cast_fp16)[name = string("transpose_361")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_96, y = transpose_97)[name = string("matrix_ac_1_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_3_begin_0 = const()[name = string("matrix_bd_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_3_end_0 = const()[name = string("matrix_bd_3_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_3_end_mask_0 = const()[name = string("matrix_bd_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1_cast_fp16)[name = string("matrix_bd_3_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_538_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = string("op_538_cast_fp16")];
+            fp16 _inversed_scores_1_y_0_to_fp16 = const()[name = string("_inversed_scores_1_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_1_cast_fp16 = mul(x = var_538_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = string("_inversed_scores_1_cast_fp16")];
+            tensor<int32, [1]> mask_11_axes_0 = const()[name = string("mask_11_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 28, 70]> mask_11 = expand_dims(axes = mask_11_axes_0, x = mask_9)[name = string("mask_11")];
+            fp16 var_45_to_fp16 = const()[name = string("op_45_to_fp16"), val = fp16(-0x1.388p+13)];
+            tensor<fp16, [1, 8, 28, 70]> scores_3_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_11)[name = string("scores_3_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_544_cast_fp16 = softmax(axis = var_59, x = scores_3_cast_fp16)[name = string("op_544_cast_fp16")];
+            fp16 var_44_to_fp16 = const()[name = string("op_44_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 70]> input_41_cast_fp16 = select(a = var_44_to_fp16, b = var_544_cast_fp16, cond = mask_11)[name = string("input_41_cast_fp16")];
+            bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)];
+            bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_1_cast_fp16)[name = string("transpose_359")];
+            tensor<fp16, [1, 8, 28, 128]> x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41_cast_fp16, y = value_9_cast_fp16)[name = string("x_13_cast_fp16")];
+            tensor<int32, [4]> var_548_perm_0 = const()[name = string("op_548_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_549 = const()[name = string("op_549"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_548_cast_fp16 = transpose(perm = var_548_perm_0, x = x_13_cast_fp16)[name = string("transpose_358")];
+            tensor<fp16, [1, 28, 1024]> input_43_cast_fp16 = reshape(shape = var_549, x = var_548_cast_fp16)[name = string("input_43_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16331136))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17379776))))[name = string("encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17381888)))];
+            tensor<fp16, [1, 28, 1024]> linear_7_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_7_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = string("input_47_cast_fp16")];
+            tensor<int32, [1]> x_17_axes_0 = const()[name = string("x_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17384000)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17386112)))];
+            tensor<fp16, [1, 28, 1024]> x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = string("x_17_cast_fp16")];
+            tensor<int32, [3]> input_49_perm_0 = const()[name = string("input_49_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_51_pad_type_0 = const()[name = string("input_51_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_51_strides_0 = const()[name = string("input_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_51_pad_0 = const()[name = string("input_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_51_dilations_0 = const()[name = string("input_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_51_groups_0 = const()[name = string("input_51_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17388224))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19485440))))[name = string("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = string("transpose_357")];
+            tensor<fp16, [1, 2048, 28]> input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("input_51_cast_fp16")];
+            int32 x_19_split_num_splits_0 = const()[name = string("x_19_split_num_splits_0"), val = int32(2)];
+            int32 x_19_split_axis_0 = const()[name = string("x_19_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_19_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51_cast_fp16)[name = string("x_19_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = string("x_19_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = string("x_19_cast_fp16")];
+            tensor<int32, [1]> var_575_axes_0 = const()[name = string("op_575_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 28]> var_575 = expand_dims(axes = var_575_axes_0, x = pad_mask)[name = string("op_575")];
+            tensor<fp16, [1, 1024, 28]> input_53_cast_fp16 = select(a = var_44_to_fp16, b = x_19_cast_fp16, cond = var_575)[name = string("input_53_cast_fp16")];
+            bool new_x_3_interleave_0 = const()[name = string("new_x_3_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_3_cast_fp16 = concat(axis = var_59, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = string("new_x_3_cast_fp16")];
+            tensor<int32, [3]> var_588_begin_0 = const()[name = string("op_588_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_588_end_0 = const()[name = string("op_588_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_588_end_mask_0 = const()[name = string("op_588_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_588_cast_fp16 = slice_by_index(begin = var_588_begin_0, end = var_588_end_0, end_mask = var_588_end_mask_0, x = new_x_3_cast_fp16)[name = string("op_588_cast_fp16")];
+            string x_21_pad_type_0 = const()[name = string("x_21_pad_type_0"), val = string("valid")];
+            int32 x_21_groups_0 = const()[name = string("x_21_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_21_strides_0 = const()[name = string("x_21_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_21_pad_0 = const()[name = string("x_21_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_21_dilations_0 = const()[name = string("x_21_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19489600))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19498880))))[name = string("encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_3_cast_fp16)[name = string("x_21_cast_fp16")];
+            tensor<int32, [3]> input_55_perm_0 = const()[name = string("input_55_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_23_axes_0 = const()[name = string("x_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19500992)))];
+            tensor<fp16, [1024]> encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19503104)))];
+            tensor<fp16, [1, 28, 1024]> input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = string("transpose_356")];
+            tensor<fp16, [1, 28, 1024]> x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = string("x_23_cast_fp16")];
+            tensor<int32, [3]> input_57_perm_0 = const()[name = string("input_57_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = string("transpose_355")];
+            tensor<fp16, [1, 1024, 28]> input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = string("input_59_cast_fp16")];
+            string x_25_pad_type_0 = const()[name = string("x_25_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_25_strides_0 = const()[name = string("x_25_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_25_pad_0 = const()[name = string("x_25_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_25_dilations_0 = const()[name = string("x_25_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_25_groups_0 = const()[name = string("x_25_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19505216))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20553856))))[name = string("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("x_25_cast_fp16")];
+            tensor<int32, [3]> input_61_perm_0 = const()[name = string("input_61_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = string("transpose_354")];
+            tensor<fp16, [1, 28, 1024]> input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = string("input_63_cast_fp16")];
+            tensor<int32, [1]> input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20555968)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20558080)))];
+            tensor<fp16, [1, 28, 1024]> input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20560192))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24754560))))[name = string("encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_0_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24762816)))];
+            tensor<fp16, [1, 28, 4096]> linear_8_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_8_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = string("input_69_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24771072))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28965440))))[name = string("encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28967552)))];
+            tensor<fp16, [1, 28, 1024]> linear_9_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_9_cast_fp16")];
+            fp16 var_631_to_fp16 = const()[name = string("op_631_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_632_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_631_to_fp16)[name = string("op_632_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_632_cast_fp16)[name = string("input_75_cast_fp16")];
+            tensor<int32, [1]> input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28969664)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28971776)))];
+            tensor<fp16, [1, 28, 1024]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")];
+            tensor<int32, [4]> cache_5_begin_0 = const()[name = string("cache_5_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_5_end_0 = const()[name = string("cache_5_end_0"), val = tensor<int32, [4]>([2, 1, 42, 1024])];
+            tensor<bool, [4]> cache_5_end_mask_0 = const()[name = string("cache_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_5_squeeze_mask_0 = const()[name = string("cache_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_5_cast_fp16 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_5_cast_fp16")];
+            tensor<int32, [4]> cache_7_begin_0 = const()[name = string("cache_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_7_end_0 = const()[name = string("cache_7_end_0"), val = tensor<int32, [4]>([2, 1, 1024, 8])];
+            tensor<bool, [4]> cache_7_end_mask_0 = const()[name = string("cache_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_7_squeeze_mask_0 = const()[name = string("cache_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_7_cast_fp16 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_7_cast_fp16")];
+            tensor<int32, [1]> input_79_axes_0 = const()[name = string("input_79_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28973888)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28976000)))];
+            tensor<fp16, [1, 28, 1024]> input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = string("input_79_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28978112))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33172480))))[name = string("encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_1_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33180736)))];
+            tensor<fp16, [1, 28, 4096]> linear_10_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_10_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = string("input_83_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33188992))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37383360))))[name = string("encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37385472)))];
+            tensor<fp16, [1, 28, 1024]> linear_11_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_11_cast_fp16")];
+            fp16 var_668_to_fp16 = const()[name = string("op_668_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_669_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_668_to_fp16)[name = string("op_669_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_669_cast_fp16)[name = string("input_89_cast_fp16")];
+            tensor<int32, [1]> key_3_axes_0 = const()[name = string("key_3_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37387584)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37389696)))];
+            tensor<fp16, [1, 28, 1024]> key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = string("key_3_cast_fp16")];
+            bool input_91_interleave_0 = const()[name = string("input_91_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_91_cast_fp16 = concat(axis = var_68, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = string("input_91_cast_fp16")];
+            tensor<int32, [3]> var_691_begin_0 = const()[name = string("op_691_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_691_end_0 = const()[name = string("op_691_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_691_end_mask_0 = const()[name = string("op_691_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_691_cast_fp16 = slice_by_index(begin = var_691_begin_0, end = var_691_end_0, end_mask = var_691_end_mask_0, x = cache_5_cast_fp16)[name = string("op_691_cast_fp16")];
+            bool var_697_interleave_0 = const()[name = string("op_697_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_697_cast_fp16 = concat(axis = var_68, interleave = var_697_interleave_0, values = (var_691_cast_fp16, key_3_cast_fp16))[name = string("op_697_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37391808))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38440448))))[name = string("encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38442560)))];
+            tensor<fp16, [1, 28, 1024]> linear_12_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized, x = key_3_cast_fp16)[name = string("linear_12_cast_fp16")];
+            tensor<int32, [4]> var_702 = const()[name = string("op_702"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_7_cast_fp16 = reshape(shape = var_702, x = linear_12_cast_fp16)[name = string("q_7_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38444672))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39493312))))[name = string("encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39495424)))];
+            tensor<fp16, [1, 70, 1024]> linear_13_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_13_cast_fp16")];
+            tensor<int32, [4]> var_707 = const()[name = string("op_707"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_5_cast_fp16 = reshape(shape = var_707, x = linear_13_cast_fp16)[name = string("k_5_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39497536))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40546176))))[name = string("encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40548288)))];
+            tensor<fp16, [1, 70, 1024]> linear_14_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_14_cast_fp16")];
+            tensor<int32, [4]> var_712 = const()[name = string("op_712"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_3_cast_fp16 = reshape(shape = var_712, x = linear_14_cast_fp16)[name = string("v_3_cast_fp16")];
+            tensor<int32, [4]> value_11_perm_0 = const()[name = string("value_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40550400)))];
+            tensor<fp16, [1, 28, 8, 128]> var_725_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = string("op_725_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40552512)))];
+            tensor<fp16, [1, 28, 8, 128]> var_727_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = string("op_727_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_3_perm_0 = const()[name = string("q_with_bias_v_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_33_transpose_x_0 = const()[name = string("x_33_transpose_x_0"), val = bool(false)];
+            bool x_33_transpose_y_0 = const()[name = string("x_33_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_729_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40554624))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40697024))))[name = string("op_729_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_727_cast_fp16)[name = string("transpose_353")];
+            tensor<fp16, [1, 8, 28, 139]> x_33_cast_fp16 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = op_729_to_fp16_quantized)[name = string("x_33_cast_fp16")];
+            tensor<int32, [8]> x_35_pad_0 = const()[name = string("x_35_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_35_mode_0 = const()[name = string("x_35_mode_0"), val = string("constant")];
+            fp16 const_92_to_fp16 = const()[name = string("const_92_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_35_cast_fp16 = pad(constant_val = const_92_to_fp16, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33_cast_fp16)[name = string("x_35_cast_fp16")];
+            tensor<int32, [4]> var_737 = const()[name = string("op_737"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_37_cast_fp16 = reshape(shape = var_737, x = x_35_cast_fp16)[name = string("x_37_cast_fp16")];
+            tensor<int32, [4]> var_741_begin_0 = const()[name = string("op_741_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_741_end_0 = const()[name = string("op_741_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_741_end_mask_0 = const()[name = string("op_741_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_741_cast_fp16 = slice_by_index(begin = var_741_begin_0, end = var_741_end_0, end_mask = var_741_end_mask_0, x = x_37_cast_fp16)[name = string("op_741_cast_fp16")];
+            tensor<int32, [4]> var_742 = const()[name = string("op_742"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_5_cast_fp16 = reshape(shape = var_742, x = var_741_cast_fp16)[name = string("matrix_bd_5_cast_fp16")];
+            bool matrix_ac_3_transpose_x_0 = const()[name = string("matrix_ac_3_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_3_transpose_y_0 = const()[name = string("matrix_ac_3_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_99 = transpose(perm = transpose_99_perm_0, x = k_5_cast_fp16)[name = string("transpose_351")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_98 = transpose(perm = transpose_98_perm_0, x = var_725_cast_fp16)[name = string("transpose_352")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_98, y = transpose_99)[name = string("matrix_ac_3_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_7_begin_0 = const()[name = string("matrix_bd_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_7_end_0 = const()[name = string("matrix_bd_7_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_7_end_mask_0 = const()[name = string("matrix_bd_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5_cast_fp16)[name = string("matrix_bd_7_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_751_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = string("op_751_cast_fp16")];
+            fp16 _inversed_scores_5_y_0_to_fp16 = const()[name = string("_inversed_scores_5_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_5_cast_fp16 = mul(x = var_751_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = string("_inversed_scores_5_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_7_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_11)[name = string("scores_7_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_757_cast_fp16 = softmax(axis = var_59, x = scores_7_cast_fp16)[name = string("op_757_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_93_cast_fp16 = select(a = var_44_to_fp16, b = var_757_cast_fp16, cond = mask_11)[name = string("input_93_cast_fp16")];
+            bool x_39_transpose_x_0 = const()[name = string("x_39_transpose_x_0"), val = bool(false)];
+            bool x_39_transpose_y_0 = const()[name = string("x_39_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_3_cast_fp16)[name = string("transpose_350")];
+            tensor<fp16, [1, 8, 28, 128]> x_39_cast_fp16 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93_cast_fp16, y = value_11_cast_fp16)[name = string("x_39_cast_fp16")];
+            tensor<int32, [4]> var_761_perm_0 = const()[name = string("op_761_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_762 = const()[name = string("op_762"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_761_cast_fp16 = transpose(perm = var_761_perm_0, x = x_39_cast_fp16)[name = string("transpose_349")];
+            tensor<fp16, [1, 28, 1024]> input_95_cast_fp16 = reshape(shape = var_762, x = var_761_cast_fp16)[name = string("input_95_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40697408))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41746048))))[name = string("encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41748160)))];
+            tensor<fp16, [1, 28, 1024]> linear_16_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_16_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = string("input_99_cast_fp16")];
+            tensor<int32, [1]> x_43_axes_0 = const()[name = string("x_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41750272)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41752384)))];
+            tensor<fp16, [1, 28, 1024]> x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = string("x_43_cast_fp16")];
+            tensor<int32, [3]> input_101_perm_0 = const()[name = string("input_101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_103_pad_type_0 = const()[name = string("input_103_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_103_strides_0 = const()[name = string("input_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_103_pad_0 = const()[name = string("input_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_103_dilations_0 = const()[name = string("input_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_103_groups_0 = const()[name = string("input_103_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41754496))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43851712))))[name = string("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = string("transpose_348")];
+            tensor<fp16, [1, 2048, 28]> input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")];
+            int32 x_45_split_num_splits_0 = const()[name = string("x_45_split_num_splits_0"), val = int32(2)];
+            int32 x_45_split_axis_0 = const()[name = string("x_45_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_45_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_45_split_cast_fp16_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103_cast_fp16)[name = string("x_45_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = string("x_45_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = string("x_45_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_105_cast_fp16 = select(a = var_44_to_fp16, b = x_45_cast_fp16, cond = var_575)[name = string("input_105_cast_fp16")];
+            bool new_x_7_interleave_0 = const()[name = string("new_x_7_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_7_cast_fp16 = concat(axis = var_59, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = string("new_x_7_cast_fp16")];
+            tensor<int32, [3]> var_801_begin_0 = const()[name = string("op_801_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_801_end_0 = const()[name = string("op_801_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_801_end_mask_0 = const()[name = string("op_801_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_801_cast_fp16 = slice_by_index(begin = var_801_begin_0, end = var_801_end_0, end_mask = var_801_end_mask_0, x = new_x_7_cast_fp16)[name = string("op_801_cast_fp16")];
+            string x_47_pad_type_0 = const()[name = string("x_47_pad_type_0"), val = string("valid")];
+            int32 x_47_groups_0 = const()[name = string("x_47_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_47_strides_0 = const()[name = string("x_47_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_47_pad_0 = const()[name = string("x_47_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_47_dilations_0 = const()[name = string("x_47_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43855872))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43865152))))[name = string("encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_7_cast_fp16)[name = string("x_47_cast_fp16")];
+            tensor<int32, [3]> input_107_perm_0 = const()[name = string("input_107_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_49_axes_0 = const()[name = string("x_49_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43867264)))];
+            tensor<fp16, [1024]> encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43869376)))];
+            tensor<fp16, [1, 28, 1024]> input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = string("transpose_347")];
+            tensor<fp16, [1, 28, 1024]> x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = string("x_49_cast_fp16")];
+            tensor<int32, [3]> input_109_perm_0 = const()[name = string("input_109_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = string("transpose_346")];
+            tensor<fp16, [1, 1024, 28]> input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = string("input_111_cast_fp16")];
+            string x_51_pad_type_0 = const()[name = string("x_51_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_51_strides_0 = const()[name = string("x_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_51_pad_0 = const()[name = string("x_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_51_dilations_0 = const()[name = string("x_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_51_groups_0 = const()[name = string("x_51_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43871488))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44920128))))[name = string("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("x_51_cast_fp16")];
+            tensor<int32, [3]> input_113_perm_0 = const()[name = string("input_113_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = string("transpose_345")];
+            tensor<fp16, [1, 28, 1024]> input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = string("input_115_cast_fp16")];
+            tensor<int32, [1]> input_117_axes_0 = const()[name = string("input_117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44922240)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44924352)))];
+            tensor<fp16, [1, 28, 1024]> input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = string("input_117_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44926464))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49120832))))[name = string("encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_1_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49129088)))];
+            tensor<fp16, [1, 28, 4096]> linear_17_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_17_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = string("input_121_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49137344))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53331712))))[name = string("encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53333824)))];
+            tensor<fp16, [1, 28, 1024]> linear_18_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_18_cast_fp16")];
+            fp16 var_844_to_fp16 = const()[name = string("op_844_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_845_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_844_to_fp16)[name = string("op_845_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_845_cast_fp16)[name = string("input_127_cast_fp16")];
+            tensor<int32, [1]> input_129_axes_0 = const()[name = string("input_129_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53335936)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53338048)))];
+            tensor<fp16, [1, 28, 1024]> input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = string("input_129_cast_fp16")];
+            tensor<int32, [4]> cache_9_begin_0 = const()[name = string("cache_9_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_9_end_0 = const()[name = string("cache_9_end_0"), val = tensor<int32, [4]>([3, 1, 42, 1024])];
+            tensor<bool, [4]> cache_9_end_mask_0 = const()[name = string("cache_9_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_9_squeeze_mask_0 = const()[name = string("cache_9_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_9_cast_fp16 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_9_cast_fp16")];
+            tensor<int32, [4]> cache_11_begin_0 = const()[name = string("cache_11_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_11_end_0 = const()[name = string("cache_11_end_0"), val = tensor<int32, [4]>([3, 1, 1024, 8])];
+            tensor<bool, [4]> cache_11_end_mask_0 = const()[name = string("cache_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_11_squeeze_mask_0 = const()[name = string("cache_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_11_cast_fp16 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_11_cast_fp16")];
+            tensor<int32, [1]> input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53340160)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53342272)))];
+            tensor<fp16, [1, 28, 1024]> input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53344384))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57538752))))[name = string("encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_2_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57547008)))];
+            tensor<fp16, [1, 28, 4096]> linear_19_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_19_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = string("input_135_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57555264))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61749632))))[name = string("encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61751744)))];
+            tensor<fp16, [1, 28, 1024]> linear_20_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_20_cast_fp16")];
+            fp16 var_881_to_fp16 = const()[name = string("op_881_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_882_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_881_to_fp16)[name = string("op_882_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_882_cast_fp16)[name = string("input_141_cast_fp16")];
+            tensor<int32, [1]> key_5_axes_0 = const()[name = string("key_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61753856)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61755968)))];
+            tensor<fp16, [1, 28, 1024]> key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = string("key_5_cast_fp16")];
+            bool input_143_interleave_0 = const()[name = string("input_143_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_143_cast_fp16 = concat(axis = var_68, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = string("input_143_cast_fp16")];
+            tensor<int32, [3]> var_904_begin_0 = const()[name = string("op_904_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_904_end_0 = const()[name = string("op_904_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_904_end_mask_0 = const()[name = string("op_904_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_904_cast_fp16 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = cache_9_cast_fp16)[name = string("op_904_cast_fp16")];
+            bool var_910_interleave_0 = const()[name = string("op_910_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_910_cast_fp16 = concat(axis = var_68, interleave = var_910_interleave_0, values = (var_904_cast_fp16, key_5_cast_fp16))[name = string("op_910_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61758080))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62806720))))[name = string("encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62808832)))];
+            tensor<fp16, [1, 28, 1024]> linear_21_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized, x = key_5_cast_fp16)[name = string("linear_21_cast_fp16")];
+            tensor<int32, [4]> var_915 = const()[name = string("op_915"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_13_cast_fp16 = reshape(shape = var_915, x = linear_21_cast_fp16)[name = string("q_13_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62810944))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63859584))))[name = string("encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63861696)))];
+            tensor<fp16, [1, 70, 1024]> linear_22_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_22_cast_fp16")];
+            tensor<int32, [4]> var_920 = const()[name = string("op_920"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_9_cast_fp16 = reshape(shape = var_920, x = linear_22_cast_fp16)[name = string("k_9_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63863808))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64912448))))[name = string("encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64914560)))];
+            tensor<fp16, [1, 70, 1024]> linear_23_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_23_cast_fp16")];
+            tensor<int32, [4]> var_925 = const()[name = string("op_925"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_5_cast_fp16 = reshape(shape = var_925, x = linear_23_cast_fp16)[name = string("v_5_cast_fp16")];
+            tensor<int32, [4]> value_13_perm_0 = const()[name = string("value_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64916672)))];
+            tensor<fp16, [1, 28, 8, 128]> var_938_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = string("op_938_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64918784)))];
+            tensor<fp16, [1, 28, 8, 128]> var_940_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = string("op_940_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_5_perm_0 = const()[name = string("q_with_bias_v_5_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_59_transpose_x_0 = const()[name = string("x_59_transpose_x_0"), val = bool(false)];
+            bool x_59_transpose_y_0 = const()[name = string("x_59_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_942_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64920896))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65063296))))[name = string("op_942_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_940_cast_fp16)[name = string("transpose_344")];
+            tensor<fp16, [1, 8, 28, 139]> x_59_cast_fp16 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = op_942_to_fp16_quantized)[name = string("x_59_cast_fp16")];
+            tensor<int32, [8]> x_61_pad_0 = const()[name = string("x_61_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_61_mode_0 = const()[name = string("x_61_mode_0"), val = string("constant")];
+            fp16 const_105_to_fp16 = const()[name = string("const_105_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_61_cast_fp16 = pad(constant_val = const_105_to_fp16, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59_cast_fp16)[name = string("x_61_cast_fp16")];
+            tensor<int32, [4]> var_950 = const()[name = string("op_950"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_63_cast_fp16 = reshape(shape = var_950, x = x_61_cast_fp16)[name = string("x_63_cast_fp16")];
+            tensor<int32, [4]> var_954_begin_0 = const()[name = string("op_954_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_954_end_0 = const()[name = string("op_954_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_954_end_mask_0 = const()[name = string("op_954_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_954_cast_fp16 = slice_by_index(begin = var_954_begin_0, end = var_954_end_0, end_mask = var_954_end_mask_0, x = x_63_cast_fp16)[name = string("op_954_cast_fp16")];
+            tensor<int32, [4]> var_955 = const()[name = string("op_955"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_9_cast_fp16 = reshape(shape = var_955, x = var_954_cast_fp16)[name = string("matrix_bd_9_cast_fp16")];
+            bool matrix_ac_5_transpose_x_0 = const()[name = string("matrix_ac_5_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_5_transpose_y_0 = const()[name = string("matrix_ac_5_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_101 = transpose(perm = transpose_101_perm_0, x = k_9_cast_fp16)[name = string("transpose_342")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_100 = transpose(perm = transpose_100_perm_0, x = var_938_cast_fp16)[name = string("transpose_343")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_100, y = transpose_101)[name = string("matrix_ac_5_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_11_begin_0 = const()[name = string("matrix_bd_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_11_end_0 = const()[name = string("matrix_bd_11_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_11_end_mask_0 = const()[name = string("matrix_bd_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9_cast_fp16)[name = string("matrix_bd_11_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_964_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = string("op_964_cast_fp16")];
+            fp16 _inversed_scores_9_y_0_to_fp16 = const()[name = string("_inversed_scores_9_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_9_cast_fp16 = mul(x = var_964_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = string("_inversed_scores_9_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_11_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_11)[name = string("scores_11_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_970_cast_fp16 = softmax(axis = var_59, x = scores_11_cast_fp16)[name = string("op_970_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_145_cast_fp16 = select(a = var_44_to_fp16, b = var_970_cast_fp16, cond = mask_11)[name = string("input_145_cast_fp16")];
+            bool x_65_transpose_x_0 = const()[name = string("x_65_transpose_x_0"), val = bool(false)];
+            bool x_65_transpose_y_0 = const()[name = string("x_65_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_5_cast_fp16)[name = string("transpose_341")];
+            tensor<fp16, [1, 8, 28, 128]> x_65_cast_fp16 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145_cast_fp16, y = value_13_cast_fp16)[name = string("x_65_cast_fp16")];
+            tensor<int32, [4]> var_974_perm_0 = const()[name = string("op_974_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_975 = const()[name = string("op_975"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_974_cast_fp16 = transpose(perm = var_974_perm_0, x = x_65_cast_fp16)[name = string("transpose_340")];
+            tensor<fp16, [1, 28, 1024]> input_147_cast_fp16 = reshape(shape = var_975, x = var_974_cast_fp16)[name = string("input_147_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65063680))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65850176))))[name = string("encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65850368)))];
+            tensor<fp16, [1, 28, 1024]> linear_25_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = string("linear_25_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = string("input_151_cast_fp16")];
+            tensor<int32, [1]> x_69_axes_0 = const()[name = string("x_69_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65852480)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65854592)))];
+            tensor<fp16, [1, 28, 1024]> x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = string("x_69_cast_fp16")];
+            tensor<int32, [3]> input_153_perm_0 = const()[name = string("input_153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_155_pad_type_0 = const()[name = string("input_155_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_155_strides_0 = const()[name = string("input_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_155_pad_0 = const()[name = string("input_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_155_dilations_0 = const()[name = string("input_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_155_groups_0 = const()[name = string("input_155_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65856704))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67953920))))[name = string("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = string("transpose_339")];
+            tensor<fp16, [1, 2048, 28]> input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")];
+            int32 x_71_split_num_splits_0 = const()[name = string("x_71_split_num_splits_0"), val = int32(2)];
+            int32 x_71_split_axis_0 = const()[name = string("x_71_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_71_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_71_split_cast_fp16_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155_cast_fp16)[name = string("x_71_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = string("x_71_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = string("x_71_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_157_cast_fp16 = select(a = var_44_to_fp16, b = x_71_cast_fp16, cond = var_575)[name = string("input_157_cast_fp16")];
+            bool new_x_11_interleave_0 = const()[name = string("new_x_11_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_11_cast_fp16 = concat(axis = var_59, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = string("new_x_11_cast_fp16")];
+            tensor<int32, [3]> var_1014_begin_0 = const()[name = string("op_1014_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_1014_end_0 = const()[name = string("op_1014_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_1014_end_mask_0 = const()[name = string("op_1014_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1014_cast_fp16 = slice_by_index(begin = var_1014_begin_0, end = var_1014_end_0, end_mask = var_1014_end_mask_0, x = new_x_11_cast_fp16)[name = string("op_1014_cast_fp16")];
+            string x_73_pad_type_0 = const()[name = string("x_73_pad_type_0"), val = string("valid")];
+            int32 x_73_groups_0 = const()[name = string("x_73_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_73_strides_0 = const()[name = string("x_73_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_73_pad_0 = const()[name = string("x_73_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_73_dilations_0 = const()[name = string("x_73_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67958080))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67967360))))[name = string("encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_73_cast_fp16 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_11_cast_fp16)[name = string("x_73_cast_fp16")];
+            tensor<int32, [3]> input_159_perm_0 = const()[name = string("input_159_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_75_axes_0 = const()[name = string("x_75_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67969472)))];
+            tensor<fp16, [1024]> encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67971584)))];
+            tensor<fp16, [1, 28, 1024]> input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = string("transpose_338")];
+            tensor<fp16, [1, 28, 1024]> x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = string("x_75_cast_fp16")];
+            tensor<int32, [3]> input_161_perm_0 = const()[name = string("input_161_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = string("transpose_337")];
+            tensor<fp16, [1, 1024, 28]> input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = string("input_163_cast_fp16")];
+            string x_77_pad_type_0 = const()[name = string("x_77_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_77_strides_0 = const()[name = string("x_77_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_77_pad_0 = const()[name = string("x_77_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_77_dilations_0 = const()[name = string("x_77_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_77_groups_0 = const()[name = string("x_77_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67973696))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69022336))))[name = string("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("x_77_cast_fp16")];
+            tensor<int32, [3]> input_165_perm_0 = const()[name = string("input_165_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = string("transpose_336")];
+            tensor<fp16, [1, 28, 1024]> input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = string("input_167_cast_fp16")];
+            tensor<int32, [1]> input_169_axes_0 = const()[name = string("input_169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69024448)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69026560)))];
+            tensor<fp16, [1, 28, 1024]> input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = string("input_169_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69028672))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72174464))))[name = string("encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_2_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72174656)))];
+            tensor<fp16, [1, 28, 4096]> linear_26_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = string("linear_26_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = string("input_173_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72182912))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75328704))))[name = string("encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_2_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75328896)))];
+            tensor<fp16, [1, 28, 1024]> linear_27_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized, x = input_173_cast_fp16)[name = string("linear_27_cast_fp16")];
+            fp16 var_1057_to_fp16 = const()[name = string("op_1057_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_1058_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1057_to_fp16)[name = string("op_1058_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_1058_cast_fp16)[name = string("input_179_cast_fp16")];
+            tensor<int32, [1]> input_181_axes_0 = const()[name = string("input_181_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75331008)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75333120)))];
+            tensor<fp16, [1, 28, 1024]> input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = string("input_181_cast_fp16")];
+            tensor<int32, [4]> cache_13_begin_0 = const()[name = string("cache_13_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_13_end_0 = const()[name = string("cache_13_end_0"), val = tensor<int32, [4]>([4, 1, 42, 1024])];
+            tensor<bool, [4]> cache_13_end_mask_0 = const()[name = string("cache_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_13_squeeze_mask_0 = const()[name = string("cache_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_13_cast_fp16 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_13_cast_fp16")];
+            tensor<int32, [4]> cache_15_begin_0 = const()[name = string("cache_15_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_15_end_0 = const()[name = string("cache_15_end_0"), val = tensor<int32, [4]>([4, 1, 1024, 8])];
+            tensor<bool, [4]> cache_15_end_mask_0 = const()[name = string("cache_15_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_15_squeeze_mask_0 = const()[name = string("cache_15_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_15_cast_fp16 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_15_cast_fp16")];
+            tensor<int32, [1]> input_183_axes_0 = const()[name = string("input_183_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75335232)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75337344)))];
+            tensor<fp16, [1, 28, 1024]> input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = string("input_183_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75339456))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78485248))))[name = string("encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_3_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78485440)))];
+            tensor<fp16, [1, 28, 4096]> linear_28_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = string("linear_28_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = string("input_187_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78493696))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81639488))))[name = string("encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81639680)))];
+            tensor<fp16, [1, 28, 1024]> linear_29_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = string("linear_29_cast_fp16")];
+            fp16 var_1094_to_fp16 = const()[name = string("op_1094_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_1095_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_1094_to_fp16)[name = string("op_1095_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_1095_cast_fp16)[name = string("input_193_cast_fp16")];
+            tensor<int32, [1]> key_7_axes_0 = const()[name = string("key_7_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81641792)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81643904)))];
+            tensor<fp16, [1, 28, 1024]> key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = string("key_7_cast_fp16")];
+            bool input_195_interleave_0 = const()[name = string("input_195_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_195_cast_fp16 = concat(axis = var_68, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = string("input_195_cast_fp16")];
+            tensor<int32, [3]> var_1117_begin_0 = const()[name = string("op_1117_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_1117_end_0 = const()[name = string("op_1117_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1117_end_mask_0 = const()[name = string("op_1117_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_1117_cast_fp16 = slice_by_index(begin = var_1117_begin_0, end = var_1117_end_0, end_mask = var_1117_end_mask_0, x = cache_13_cast_fp16)[name = string("op_1117_cast_fp16")];
+            bool var_1123_interleave_0 = const()[name = string("op_1123_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1123_cast_fp16 = concat(axis = var_68, interleave = var_1123_interleave_0, values = (var_1117_cast_fp16, key_7_cast_fp16))[name = string("op_1123_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81646016))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82432512))))[name = string("encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82432704)))];
+            tensor<fp16, [1, 28, 1024]> linear_30_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized, x = key_7_cast_fp16)[name = string("linear_30_cast_fp16")];
+            tensor<int32, [4]> var_1128 = const()[name = string("op_1128"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_19_cast_fp16 = reshape(shape = var_1128, x = linear_30_cast_fp16)[name = string("q_19_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82434816))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83221312))))[name = string("encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83221504)))];
+            tensor<fp16, [1, 70, 1024]> linear_31_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_31_cast_fp16")];
+            tensor<int32, [4]> var_1133 = const()[name = string("op_1133"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_13_cast_fp16 = reshape(shape = var_1133, x = linear_31_cast_fp16)[name = string("k_13_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83223616))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84010112))))[name = string("encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84010304)))];
+            tensor<fp16, [1, 70, 1024]> linear_32_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_32_cast_fp16")];
+            tensor<int32, [4]> var_1138 = const()[name = string("op_1138"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_7_cast_fp16 = reshape(shape = var_1138, x = linear_32_cast_fp16)[name = string("v_7_cast_fp16")];
+            tensor<int32, [4]> value_15_perm_0 = const()[name = string("value_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84012416)))];
+            tensor<fp16, [1, 28, 8, 128]> var_1151_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = string("op_1151_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84014528)))];
+            tensor<fp16, [1, 28, 8, 128]> var_1153_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = string("op_1153_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_7_perm_0 = const()[name = string("q_with_bias_v_7_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_85_transpose_x_0 = const()[name = string("x_85_transpose_x_0"), val = bool(false)];
+            bool x_85_transpose_y_0 = const()[name = string("x_85_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_1155_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84016640))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84159040))))[name = string("op_1155_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_1153_cast_fp16)[name = string("transpose_335")];
+            tensor<fp16, [1, 8, 28, 139]> x_85_cast_fp16 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = op_1155_to_fp16_quantized)[name = string("x_85_cast_fp16")];
+            tensor<int32, [8]> x_87_pad_0 = const()[name = string("x_87_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("constant")];
+            fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_87_cast_fp16 = pad(constant_val = const_118_to_fp16, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85_cast_fp16)[name = string("x_87_cast_fp16")];
+            tensor<int32, [4]> var_1163 = const()[name = string("op_1163"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_89_cast_fp16 = reshape(shape = var_1163, x = x_87_cast_fp16)[name = string("x_89_cast_fp16")];
+            tensor<int32, [4]> var_1167_begin_0 = const()[name = string("op_1167_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1167_end_0 = const()[name = string("op_1167_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_1167_end_mask_0 = const()[name = string("op_1167_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_1167_cast_fp16 = slice_by_index(begin = var_1167_begin_0, end = var_1167_end_0, end_mask = var_1167_end_mask_0, x = x_89_cast_fp16)[name = string("op_1167_cast_fp16")];
+            tensor<int32, [4]> var_1168 = const()[name = string("op_1168"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_13_cast_fp16 = reshape(shape = var_1168, x = var_1167_cast_fp16)[name = string("matrix_bd_13_cast_fp16")];
+            bool matrix_ac_7_transpose_x_0 = const()[name = string("matrix_ac_7_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_7_transpose_y_0 = const()[name = string("matrix_ac_7_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_103 = transpose(perm = transpose_103_perm_0, x = k_13_cast_fp16)[name = string("transpose_333")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_102 = transpose(perm = transpose_102_perm_0, x = var_1151_cast_fp16)[name = string("transpose_334")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_102, y = transpose_103)[name = string("matrix_ac_7_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_15_begin_0 = const()[name = string("matrix_bd_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_15_end_0 = const()[name = string("matrix_bd_15_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_15_end_mask_0 = const()[name = string("matrix_bd_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13_cast_fp16)[name = string("matrix_bd_15_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_1177_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = string("op_1177_cast_fp16")];
+            fp16 _inversed_scores_13_y_0_to_fp16 = const()[name = string("_inversed_scores_13_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_13_cast_fp16 = mul(x = var_1177_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = string("_inversed_scores_13_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_15_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_11)[name = string("scores_15_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_1183_cast_fp16 = softmax(axis = var_59, x = scores_15_cast_fp16)[name = string("op_1183_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_197_cast_fp16 = select(a = var_44_to_fp16, b = var_1183_cast_fp16, cond = mask_11)[name = string("input_197_cast_fp16")];
+            bool x_91_transpose_x_0 = const()[name = string("x_91_transpose_x_0"), val = bool(false)];
+            bool x_91_transpose_y_0 = const()[name = string("x_91_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_7_cast_fp16)[name = string("transpose_332")];
+            tensor<fp16, [1, 8, 28, 128]> x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197_cast_fp16, y = value_15_cast_fp16)[name = string("x_91_cast_fp16")];
+            tensor<int32, [4]> var_1187_perm_0 = const()[name = string("op_1187_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1188 = const()[name = string("op_1188"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_1187_cast_fp16 = transpose(perm = var_1187_perm_0, x = x_91_cast_fp16)[name = string("transpose_331")];
+            tensor<fp16, [1, 28, 1024]> input_199_cast_fp16 = reshape(shape = var_1188, x = var_1187_cast_fp16)[name = string("input_199_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84159424))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84945920))))[name = string("encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84946112)))];
+            tensor<fp16, [1, 28, 1024]> linear_34_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = string("linear_34_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = string("input_203_cast_fp16")];
+            tensor<int32, [1]> x_95_axes_0 = const()[name = string("x_95_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84948224)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84950336)))];
+            tensor<fp16, [1, 28, 1024]> x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = string("x_95_cast_fp16")];
+            tensor<int32, [3]> input_205_perm_0 = const()[name = string("input_205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_207_pad_type_0 = const()[name = string("input_207_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_207_strides_0 = const()[name = string("input_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_207_pad_0 = const()[name = string("input_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_207_dilations_0 = const()[name = string("input_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_207_groups_0 = const()[name = string("input_207_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84952448))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87049664))))[name = string("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = string("transpose_330")];
+            tensor<fp16, [1, 2048, 28]> input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("input_207_cast_fp16")];
+            int32 x_97_split_num_splits_0 = const()[name = string("x_97_split_num_splits_0"), val = int32(2)];
+            int32 x_97_split_axis_0 = const()[name = string("x_97_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_97_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207_cast_fp16)[name = string("x_97_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = string("x_97_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = string("x_97_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_209_cast_fp16 = select(a = var_44_to_fp16, b = x_97_cast_fp16, cond = var_575)[name = string("input_209_cast_fp16")];
+            bool new_x_15_interleave_0 = const()[name = string("new_x_15_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_15_cast_fp16 = concat(axis = var_59, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = string("new_x_15_cast_fp16")];
+            tensor<int32, [3]> var_1227_begin_0 = const()[name = string("op_1227_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_1227_end_0 = const()[name = string("op_1227_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_1227_end_mask_0 = const()[name = string("op_1227_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1227_cast_fp16 = slice_by_index(begin = var_1227_begin_0, end = var_1227_end_0, end_mask = var_1227_end_mask_0, x = new_x_15_cast_fp16)[name = string("op_1227_cast_fp16")];
+            string x_99_pad_type_0 = const()[name = string("x_99_pad_type_0"), val = string("valid")];
+            int32 x_99_groups_0 = const()[name = string("x_99_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_99_strides_0 = const()[name = string("x_99_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_99_pad_0 = const()[name = string("x_99_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_99_dilations_0 = const()[name = string("x_99_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87053824))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87063104))))[name = string("encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_15_cast_fp16)[name = string("x_99_cast_fp16")];
+            tensor<int32, [3]> input_211_perm_0 = const()[name = string("input_211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_101_axes_0 = const()[name = string("x_101_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87065216)))];
+            tensor<fp16, [1024]> encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87067328)))];
+            tensor<fp16, [1, 28, 1024]> input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = string("transpose_329")];
+            tensor<fp16, [1, 28, 1024]> x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = string("x_101_cast_fp16")];
+            tensor<int32, [3]> input_213_perm_0 = const()[name = string("input_213_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = string("transpose_328")];
+            tensor<fp16, [1, 1024, 28]> input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = string("input_215_cast_fp16")];
+            string x_103_pad_type_0 = const()[name = string("x_103_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_103_strides_0 = const()[name = string("x_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_103_pad_0 = const()[name = string("x_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_103_dilations_0 = const()[name = string("x_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_103_groups_0 = const()[name = string("x_103_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87069440))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88118080))))[name = string("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("x_103_cast_fp16")];
+            tensor<int32, [3]> input_217_perm_0 = const()[name = string("input_217_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = string("transpose_327")];
+            tensor<fp16, [1, 28, 1024]> input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = string("input_219_cast_fp16")];
+            tensor<int32, [1]> input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88120192)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88122304)))];
+            tensor<fp16, [1, 28, 1024]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88124416))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91270208))))[name = string("encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_3_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91270400)))];
+            tensor<fp16, [1, 28, 4096]> linear_35_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = string("linear_35_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = string("input_225_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91278656))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94424448))))[name = string("encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94424640)))];
+            tensor<fp16, [1, 28, 1024]> linear_36_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = string("linear_36_cast_fp16")];
+            fp16 var_1270_to_fp16 = const()[name = string("op_1270_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_1271_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1270_to_fp16)[name = string("op_1271_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1271_cast_fp16)[name = string("input_231_cast_fp16")];
+            tensor<int32, [1]> input_233_axes_0 = const()[name = string("input_233_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94426752)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94428864)))];
+            tensor<fp16, [1, 28, 1024]> input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = string("input_233_cast_fp16")];
+            tensor<int32, [4]> cache_17_begin_0 = const()[name = string("cache_17_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_17_end_0 = const()[name = string("cache_17_end_0"), val = tensor<int32, [4]>([5, 1, 42, 1024])];
+            tensor<bool, [4]> cache_17_end_mask_0 = const()[name = string("cache_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_17_squeeze_mask_0 = const()[name = string("cache_17_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_17_cast_fp16 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_17_cast_fp16")];
+            tensor<int32, [4]> cache_19_begin_0 = const()[name = string("cache_19_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_19_end_0 = const()[name = string("cache_19_end_0"), val = tensor<int32, [4]>([5, 1, 1024, 8])];
+            tensor<bool, [4]> cache_19_end_mask_0 = const()[name = string("cache_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_19_squeeze_mask_0 = const()[name = string("cache_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_19_cast_fp16 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_19_cast_fp16")];
+            tensor<int32, [1]> input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94430976)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94433088)))];
+            tensor<fp16, [1, 28, 1024]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94435200))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97580992))))[name = string("encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_4_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97581184)))];
+            tensor<fp16, [1, 28, 4096]> linear_37_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized, x = input_235_cast_fp16)[name = string("linear_37_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = string("input_239_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97589440))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100735232))))[name = string("encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100735424)))];
+            tensor<fp16, [1, 28, 1024]> linear_38_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = string("linear_38_cast_fp16")];
+            fp16 var_1307_to_fp16 = const()[name = string("op_1307_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_1308_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1307_to_fp16)[name = string("op_1308_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1308_cast_fp16)[name = string("input_245_cast_fp16")];
+            tensor<int32, [1]> key_9_axes_0 = const()[name = string("key_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100737536)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100739648)))];
+            tensor<fp16, [1, 28, 1024]> key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = string("key_9_cast_fp16")];
+            bool input_247_interleave_0 = const()[name = string("input_247_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_247_cast_fp16 = concat(axis = var_68, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = string("input_247_cast_fp16")];
+            tensor<int32, [3]> var_1330_begin_0 = const()[name = string("op_1330_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_1330_end_0 = const()[name = string("op_1330_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1330_end_mask_0 = const()[name = string("op_1330_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_1330_cast_fp16 = slice_by_index(begin = var_1330_begin_0, end = var_1330_end_0, end_mask = var_1330_end_mask_0, x = cache_17_cast_fp16)[name = string("op_1330_cast_fp16")];
+            bool var_1336_interleave_0 = const()[name = string("op_1336_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1336_cast_fp16 = concat(axis = var_68, interleave = var_1336_interleave_0, values = (var_1330_cast_fp16, key_9_cast_fp16))[name = string("op_1336_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100741760))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101528256))))[name = string("encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101528448)))];
+            tensor<fp16, [1, 28, 1024]> linear_39_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized, x = key_9_cast_fp16)[name = string("linear_39_cast_fp16")];
+            tensor<int32, [4]> var_1341 = const()[name = string("op_1341"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_25_cast_fp16 = reshape(shape = var_1341, x = linear_39_cast_fp16)[name = string("q_25_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101530560))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102317056))))[name = string("encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102317248)))];
+            tensor<fp16, [1, 70, 1024]> linear_40_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_40_cast_fp16")];
+            tensor<int32, [4]> var_1346 = const()[name = string("op_1346"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_17_cast_fp16 = reshape(shape = var_1346, x = linear_40_cast_fp16)[name = string("k_17_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102319360))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103105856))))[name = string("encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103106048)))];
+            tensor<fp16, [1, 70, 1024]> linear_41_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_41_cast_fp16")];
+            tensor<int32, [4]> var_1351 = const()[name = string("op_1351"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_9_cast_fp16 = reshape(shape = var_1351, x = linear_41_cast_fp16)[name = string("v_9_cast_fp16")];
+            tensor<int32, [4]> value_17_perm_0 = const()[name = string("value_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103108160)))];
+            tensor<fp16, [1, 28, 8, 128]> var_1364_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = string("op_1364_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103110272)))];
+            tensor<fp16, [1, 28, 8, 128]> var_1366_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = string("op_1366_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_9_perm_0 = const()[name = string("q_with_bias_v_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_111_transpose_x_0 = const()[name = string("x_111_transpose_x_0"), val = bool(false)];
+            bool x_111_transpose_y_0 = const()[name = string("x_111_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_1368_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103112384))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103254784))))[name = string("op_1368_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1366_cast_fp16)[name = string("transpose_326")];
+            tensor<fp16, [1, 8, 28, 139]> x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = op_1368_to_fp16_quantized)[name = string("x_111_cast_fp16")];
+            tensor<int32, [8]> x_113_pad_0 = const()[name = string("x_113_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_113_mode_0 = const()[name = string("x_113_mode_0"), val = string("constant")];
+            fp16 const_131_to_fp16 = const()[name = string("const_131_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_113_cast_fp16 = pad(constant_val = const_131_to_fp16, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111_cast_fp16)[name = string("x_113_cast_fp16")];
+            tensor<int32, [4]> var_1376 = const()[name = string("op_1376"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_115_cast_fp16 = reshape(shape = var_1376, x = x_113_cast_fp16)[name = string("x_115_cast_fp16")];
+            tensor<int32, [4]> var_1380_begin_0 = const()[name = string("op_1380_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1380_end_0 = const()[name = string("op_1380_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_1380_end_mask_0 = const()[name = string("op_1380_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_1380_cast_fp16 = slice_by_index(begin = var_1380_begin_0, end = var_1380_end_0, end_mask = var_1380_end_mask_0, x = x_115_cast_fp16)[name = string("op_1380_cast_fp16")];
+            tensor<int32, [4]> var_1381 = const()[name = string("op_1381"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_17_cast_fp16 = reshape(shape = var_1381, x = var_1380_cast_fp16)[name = string("matrix_bd_17_cast_fp16")];
+            bool matrix_ac_9_transpose_x_0 = const()[name = string("matrix_ac_9_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_9_transpose_y_0 = const()[name = string("matrix_ac_9_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_105 = transpose(perm = transpose_105_perm_0, x = k_17_cast_fp16)[name = string("transpose_324")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1364_cast_fp16)[name = string("transpose_325")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_104, y = transpose_105)[name = string("matrix_ac_9_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_19_begin_0 = const()[name = string("matrix_bd_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_19_end_0 = const()[name = string("matrix_bd_19_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_19_end_mask_0 = const()[name = string("matrix_bd_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17_cast_fp16)[name = string("matrix_bd_19_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_1390_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = string("op_1390_cast_fp16")];
+            fp16 _inversed_scores_17_y_0_to_fp16 = const()[name = string("_inversed_scores_17_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_17_cast_fp16 = mul(x = var_1390_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = string("_inversed_scores_17_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_19_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_11)[name = string("scores_19_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_1396_cast_fp16 = softmax(axis = var_59, x = scores_19_cast_fp16)[name = string("op_1396_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_249_cast_fp16 = select(a = var_44_to_fp16, b = var_1396_cast_fp16, cond = mask_11)[name = string("input_249_cast_fp16")];
+            bool x_117_transpose_x_0 = const()[name = string("x_117_transpose_x_0"), val = bool(false)];
+            bool x_117_transpose_y_0 = const()[name = string("x_117_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_9_cast_fp16)[name = string("transpose_323")];
+            tensor<fp16, [1, 8, 28, 128]> x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249_cast_fp16, y = value_17_cast_fp16)[name = string("x_117_cast_fp16")];
+            tensor<int32, [4]> var_1400_perm_0 = const()[name = string("op_1400_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1401 = const()[name = string("op_1401"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_1400_cast_fp16 = transpose(perm = var_1400_perm_0, x = x_117_cast_fp16)[name = string("transpose_322")];
+            tensor<fp16, [1, 28, 1024]> input_251_cast_fp16 = reshape(shape = var_1401, x = var_1400_cast_fp16)[name = string("input_251_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103255168))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104041664))))[name = string("encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104041856)))];
+            tensor<fp16, [1, 28, 1024]> linear_43_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = string("linear_43_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = string("input_255_cast_fp16")];
+            tensor<int32, [1]> x_121_axes_0 = const()[name = string("x_121_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104043968)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104046080)))];
+            tensor<fp16, [1, 28, 1024]> x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = string("x_121_cast_fp16")];
+            tensor<int32, [3]> input_257_perm_0 = const()[name = string("input_257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_259_pad_type_0 = const()[name = string("input_259_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_259_strides_0 = const()[name = string("input_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_259_pad_0 = const()[name = string("input_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_259_dilations_0 = const()[name = string("input_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_259_groups_0 = const()[name = string("input_259_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104048192))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106145408))))[name = string("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = string("transpose_321")];
+            tensor<fp16, [1, 2048, 28]> input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("input_259_cast_fp16")];
+            int32 x_123_split_num_splits_0 = const()[name = string("x_123_split_num_splits_0"), val = int32(2)];
+            int32 x_123_split_axis_0 = const()[name = string("x_123_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_123_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_123_split_cast_fp16_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259_cast_fp16)[name = string("x_123_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = string("x_123_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = string("x_123_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_261_cast_fp16 = select(a = var_44_to_fp16, b = x_123_cast_fp16, cond = var_575)[name = string("input_261_cast_fp16")];
+            bool new_x_19_interleave_0 = const()[name = string("new_x_19_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_19_cast_fp16 = concat(axis = var_59, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = string("new_x_19_cast_fp16")];
+            tensor<int32, [3]> var_1440_begin_0 = const()[name = string("op_1440_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_1440_end_0 = const()[name = string("op_1440_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_1440_end_mask_0 = const()[name = string("op_1440_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1440_cast_fp16 = slice_by_index(begin = var_1440_begin_0, end = var_1440_end_0, end_mask = var_1440_end_mask_0, x = new_x_19_cast_fp16)[name = string("op_1440_cast_fp16")];
+            string x_125_pad_type_0 = const()[name = string("x_125_pad_type_0"), val = string("valid")];
+            int32 x_125_groups_0 = const()[name = string("x_125_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_125_strides_0 = const()[name = string("x_125_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_125_pad_0 = const()[name = string("x_125_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_125_dilations_0 = const()[name = string("x_125_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106149568))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106158848))))[name = string("encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_19_cast_fp16)[name = string("x_125_cast_fp16")];
+            tensor<int32, [3]> input_263_perm_0 = const()[name = string("input_263_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_127_axes_0 = const()[name = string("x_127_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106160960)))];
+            tensor<fp16, [1024]> encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106163072)))];
+            tensor<fp16, [1, 28, 1024]> input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = string("transpose_320")];
+            tensor<fp16, [1, 28, 1024]> x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = string("x_127_cast_fp16")];
+            tensor<int32, [3]> input_265_perm_0 = const()[name = string("input_265_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = string("transpose_319")];
+            tensor<fp16, [1, 1024, 28]> input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = string("input_267_cast_fp16")];
+            string x_129_pad_type_0 = const()[name = string("x_129_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_129_strides_0 = const()[name = string("x_129_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_129_pad_0 = const()[name = string("x_129_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_129_dilations_0 = const()[name = string("x_129_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_129_groups_0 = const()[name = string("x_129_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106165184))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107213824))))[name = string("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_129_cast_fp16 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("x_129_cast_fp16")];
+            tensor<int32, [3]> input_269_perm_0 = const()[name = string("input_269_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = string("transpose_318")];
+            tensor<fp16, [1, 28, 1024]> input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = string("input_271_cast_fp16")];
+            tensor<int32, [1]> input_273_axes_0 = const()[name = string("input_273_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107215936)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107218048)))];
+            tensor<fp16, [1, 28, 1024]> input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = string("input_273_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107220160))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110365952))))[name = string("encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_4_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110366144)))];
+            tensor<fp16, [1, 28, 4096]> linear_44_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized, x = input_273_cast_fp16)[name = string("linear_44_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = string("input_277_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110374400))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113520192))))[name = string("encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113520384)))];
+            tensor<fp16, [1, 28, 1024]> linear_45_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized, x = input_277_cast_fp16)[name = string("linear_45_cast_fp16")];
+            fp16 var_1483_to_fp16 = const()[name = string("op_1483_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_1484_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1483_to_fp16)[name = string("op_1484_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1484_cast_fp16)[name = string("input_283_cast_fp16")];
+            tensor<int32, [1]> input_285_axes_0 = const()[name = string("input_285_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113522496)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113524608)))];
+            tensor<fp16, [1, 28, 1024]> input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = string("input_285_cast_fp16")];
+            tensor<int32, [4]> cache_21_begin_0 = const()[name = string("cache_21_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_21_end_0 = const()[name = string("cache_21_end_0"), val = tensor<int32, [4]>([6, 1, 42, 1024])];
+            tensor<bool, [4]> cache_21_end_mask_0 = const()[name = string("cache_21_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_21_squeeze_mask_0 = const()[name = string("cache_21_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_21_cast_fp16 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_21_cast_fp16")];
+            tensor<int32, [4]> cache_23_begin_0 = const()[name = string("cache_23_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_23_end_0 = const()[name = string("cache_23_end_0"), val = tensor<int32, [4]>([6, 1, 1024, 8])];
+            tensor<bool, [4]> cache_23_end_mask_0 = const()[name = string("cache_23_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_23_squeeze_mask_0 = const()[name = string("cache_23_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_23_cast_fp16 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_23_cast_fp16")];
+            tensor<int32, [1]> input_287_axes_0 = const()[name = string("input_287_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113526720)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113528832)))];
+            tensor<fp16, [1, 28, 1024]> input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = string("input_287_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113530944))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116676736))))[name = string("encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_5_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116676928)))];
+            tensor<fp16, [1, 28, 4096]> linear_46_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = string("linear_46_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("input_291_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116685184))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119830976))))[name = string("encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119831168)))];
+            tensor<fp16, [1, 28, 1024]> linear_47_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized, x = input_291_cast_fp16)[name = string("linear_47_cast_fp16")];
+            fp16 var_1520_to_fp16 = const()[name = string("op_1520_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_1521_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1520_to_fp16)[name = string("op_1521_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1521_cast_fp16)[name = string("input_297_cast_fp16")];
+            tensor<int32, [1]> key_11_axes_0 = const()[name = string("key_11_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119833280)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119835392)))];
+            tensor<fp16, [1, 28, 1024]> key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = string("key_11_cast_fp16")];
+            bool input_299_interleave_0 = const()[name = string("input_299_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_299_cast_fp16 = concat(axis = var_68, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = string("input_299_cast_fp16")];
+            tensor<int32, [3]> var_1543_begin_0 = const()[name = string("op_1543_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_1543_end_0 = const()[name = string("op_1543_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1543_end_mask_0 = const()[name = string("op_1543_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_1543_cast_fp16 = slice_by_index(begin = var_1543_begin_0, end = var_1543_end_0, end_mask = var_1543_end_mask_0, x = cache_21_cast_fp16)[name = string("op_1543_cast_fp16")];
+            bool var_1549_interleave_0 = const()[name = string("op_1549_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1549_cast_fp16 = concat(axis = var_68, interleave = var_1549_interleave_0, values = (var_1543_cast_fp16, key_11_cast_fp16))[name = string("op_1549_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119837504))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120624000))))[name = string("encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120624192)))];
+            tensor<fp16, [1, 28, 1024]> linear_48_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized, x = key_11_cast_fp16)[name = string("linear_48_cast_fp16")];
+            tensor<int32, [4]> var_1554 = const()[name = string("op_1554"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_31_cast_fp16 = reshape(shape = var_1554, x = linear_48_cast_fp16)[name = string("q_31_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120626304))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121412800))))[name = string("encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121412992)))];
+            tensor<fp16, [1, 70, 1024]> linear_49_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_49_cast_fp16")];
+            tensor<int32, [4]> var_1559 = const()[name = string("op_1559"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_21_cast_fp16 = reshape(shape = var_1559, x = linear_49_cast_fp16)[name = string("k_21_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121415104))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122201600))))[name = string("encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122201792)))];
+            tensor<fp16, [1, 70, 1024]> linear_50_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_50_cast_fp16")];
+            tensor<int32, [4]> var_1564 = const()[name = string("op_1564"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_11_cast_fp16 = reshape(shape = var_1564, x = linear_50_cast_fp16)[name = string("v_11_cast_fp16")];
+            tensor<int32, [4]> value_19_perm_0 = const()[name = string("value_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122203904)))];
+            tensor<fp16, [1, 28, 8, 128]> var_1577_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = string("op_1577_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122206016)))];
+            tensor<fp16, [1, 28, 8, 128]> var_1579_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = string("op_1579_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_11_perm_0 = const()[name = string("q_with_bias_v_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_137_transpose_x_0 = const()[name = string("x_137_transpose_x_0"), val = bool(false)];
+            bool x_137_transpose_y_0 = const()[name = string("x_137_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_1581_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122208128))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122350528))))[name = string("op_1581_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1579_cast_fp16)[name = string("transpose_317")];
+            tensor<fp16, [1, 8, 28, 139]> x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = op_1581_to_fp16_quantized)[name = string("x_137_cast_fp16")];
+            tensor<int32, [8]> x_139_pad_0 = const()[name = string("x_139_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_139_mode_0 = const()[name = string("x_139_mode_0"), val = string("constant")];
+            fp16 const_144_to_fp16 = const()[name = string("const_144_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_139_cast_fp16 = pad(constant_val = const_144_to_fp16, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137_cast_fp16)[name = string("x_139_cast_fp16")];
+            tensor<int32, [4]> var_1589 = const()[name = string("op_1589"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_141_cast_fp16 = reshape(shape = var_1589, x = x_139_cast_fp16)[name = string("x_141_cast_fp16")];
+            tensor<int32, [4]> var_1593_begin_0 = const()[name = string("op_1593_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1593_end_0 = const()[name = string("op_1593_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_1593_end_mask_0 = const()[name = string("op_1593_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_1593_cast_fp16 = slice_by_index(begin = var_1593_begin_0, end = var_1593_end_0, end_mask = var_1593_end_mask_0, x = x_141_cast_fp16)[name = string("op_1593_cast_fp16")];
+            tensor<int32, [4]> var_1594 = const()[name = string("op_1594"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_21_cast_fp16 = reshape(shape = var_1594, x = var_1593_cast_fp16)[name = string("matrix_bd_21_cast_fp16")];
+            bool matrix_ac_11_transpose_x_0 = const()[name = string("matrix_ac_11_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_11_transpose_y_0 = const()[name = string("matrix_ac_11_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_107 = transpose(perm = transpose_107_perm_0, x = k_21_cast_fp16)[name = string("transpose_315")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_106 = transpose(perm = transpose_106_perm_0, x = var_1577_cast_fp16)[name = string("transpose_316")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_106, y = transpose_107)[name = string("matrix_ac_11_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_23_begin_0 = const()[name = string("matrix_bd_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_23_end_0 = const()[name = string("matrix_bd_23_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_23_end_mask_0 = const()[name = string("matrix_bd_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21_cast_fp16)[name = string("matrix_bd_23_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_1603_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = string("op_1603_cast_fp16")];
+            fp16 _inversed_scores_21_y_0_to_fp16 = const()[name = string("_inversed_scores_21_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_21_cast_fp16 = mul(x = var_1603_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = string("_inversed_scores_21_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_23_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_11)[name = string("scores_23_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_1609_cast_fp16 = softmax(axis = var_59, x = scores_23_cast_fp16)[name = string("op_1609_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_301_cast_fp16 = select(a = var_44_to_fp16, b = var_1609_cast_fp16, cond = mask_11)[name = string("input_301_cast_fp16")];
+            bool x_143_transpose_x_0 = const()[name = string("x_143_transpose_x_0"), val = bool(false)];
+            bool x_143_transpose_y_0 = const()[name = string("x_143_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_11_cast_fp16)[name = string("transpose_314")];
+            tensor<fp16, [1, 8, 28, 128]> x_143_cast_fp16 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301_cast_fp16, y = value_19_cast_fp16)[name = string("x_143_cast_fp16")];
+            tensor<int32, [4]> var_1613_perm_0 = const()[name = string("op_1613_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1614 = const()[name = string("op_1614"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_1613_cast_fp16 = transpose(perm = var_1613_perm_0, x = x_143_cast_fp16)[name = string("transpose_313")];
+            tensor<fp16, [1, 28, 1024]> input_303_cast_fp16 = reshape(shape = var_1614, x = var_1613_cast_fp16)[name = string("input_303_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122350912))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123137408))))[name = string("encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123137600)))];
+            tensor<fp16, [1, 28, 1024]> linear_52_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized, x = input_303_cast_fp16)[name = string("linear_52_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = string("input_307_cast_fp16")];
+            tensor<int32, [1]> x_147_axes_0 = const()[name = string("x_147_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123139712)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123141824)))];
+            tensor<fp16, [1, 28, 1024]> x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = string("x_147_cast_fp16")];
+            tensor<int32, [3]> input_309_perm_0 = const()[name = string("input_309_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_311_pad_type_0 = const()[name = string("input_311_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_311_strides_0 = const()[name = string("input_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_311_pad_0 = const()[name = string("input_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_311_dilations_0 = const()[name = string("input_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_311_groups_0 = const()[name = string("input_311_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123143936))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125241152))))[name = string("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = string("transpose_312")];
+            tensor<fp16, [1, 2048, 28]> input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_309_cast_fp16)[name = string("input_311_cast_fp16")];
+            int32 x_149_split_num_splits_0 = const()[name = string("x_149_split_num_splits_0"), val = int32(2)];
+            int32 x_149_split_axis_0 = const()[name = string("x_149_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_149_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_149_split_cast_fp16_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311_cast_fp16)[name = string("x_149_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = string("x_149_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = string("x_149_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_313_cast_fp16 = select(a = var_44_to_fp16, b = x_149_cast_fp16, cond = var_575)[name = string("input_313_cast_fp16")];
+            bool new_x_23_interleave_0 = const()[name = string("new_x_23_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_23_cast_fp16 = concat(axis = var_59, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = string("new_x_23_cast_fp16")];
+            tensor<int32, [3]> var_1653_begin_0 = const()[name = string("op_1653_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_1653_end_0 = const()[name = string("op_1653_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_1653_end_mask_0 = const()[name = string("op_1653_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1653_cast_fp16 = slice_by_index(begin = var_1653_begin_0, end = var_1653_end_0, end_mask = var_1653_end_mask_0, x = new_x_23_cast_fp16)[name = string("op_1653_cast_fp16")];
+            string x_151_pad_type_0 = const()[name = string("x_151_pad_type_0"), val = string("valid")];
+            int32 x_151_groups_0 = const()[name = string("x_151_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_151_strides_0 = const()[name = string("x_151_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_151_pad_0 = const()[name = string("x_151_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_151_dilations_0 = const()[name = string("x_151_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125245312))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125254592))))[name = string("encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_151_cast_fp16 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_23_cast_fp16)[name = string("x_151_cast_fp16")];
+            tensor<int32, [3]> input_315_perm_0 = const()[name = string("input_315_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_153_axes_0 = const()[name = string("x_153_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125256704)))];
+            tensor<fp16, [1024]> encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125258816)))];
+            tensor<fp16, [1, 28, 1024]> input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = string("transpose_311")];
+            tensor<fp16, [1, 28, 1024]> x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = string("x_153_cast_fp16")];
+            tensor<int32, [3]> input_317_perm_0 = const()[name = string("input_317_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = string("transpose_310")];
+            tensor<fp16, [1, 1024, 28]> input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = string("input_319_cast_fp16")];
+            string x_155_pad_type_0 = const()[name = string("x_155_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_155_strides_0 = const()[name = string("x_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_155_pad_0 = const()[name = string("x_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_155_dilations_0 = const()[name = string("x_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_155_groups_0 = const()[name = string("x_155_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125260928))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126309568))))[name = string("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_319_cast_fp16)[name = string("x_155_cast_fp16")];
+            tensor<int32, [3]> input_321_perm_0 = const()[name = string("input_321_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = string("transpose_309")];
+            tensor<fp16, [1, 28, 1024]> input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = string("input_323_cast_fp16")];
+            tensor<int32, [1]> input_325_axes_0 = const()[name = string("input_325_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126311680)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126313792)))];
+            tensor<fp16, [1, 28, 1024]> input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = string("input_325_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126315904))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129461696))))[name = string("encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_5_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129461888)))];
+            tensor<fp16, [1, 28, 4096]> linear_53_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized, x = input_325_cast_fp16)[name = string("linear_53_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("input_329_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129470144))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132615936))))[name = string("encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132616128)))];
+            tensor<fp16, [1, 28, 1024]> linear_54_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized, x = input_329_cast_fp16)[name = string("linear_54_cast_fp16")];
+            fp16 var_1696_to_fp16 = const()[name = string("op_1696_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_1697_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1696_to_fp16)[name = string("op_1697_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1697_cast_fp16)[name = string("input_335_cast_fp16")];
+            tensor<int32, [1]> input_337_axes_0 = const()[name = string("input_337_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132618240)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132620352)))];
+            tensor<fp16, [1, 28, 1024]> input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = string("input_337_cast_fp16")];
+            tensor<int32, [4]> cache_25_begin_0 = const()[name = string("cache_25_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_25_end_0 = const()[name = string("cache_25_end_0"), val = tensor<int32, [4]>([7, 1, 42, 1024])];
+            tensor<bool, [4]> cache_25_end_mask_0 = const()[name = string("cache_25_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_25_squeeze_mask_0 = const()[name = string("cache_25_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_25_cast_fp16 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_25_cast_fp16")];
+            tensor<int32, [4]> cache_27_begin_0 = const()[name = string("cache_27_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_27_end_0 = const()[name = string("cache_27_end_0"), val = tensor<int32, [4]>([7, 1, 1024, 8])];
+            tensor<bool, [4]> cache_27_end_mask_0 = const()[name = string("cache_27_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_27_squeeze_mask_0 = const()[name = string("cache_27_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_27_cast_fp16 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_27_cast_fp16")];
+            tensor<int32, [1]> input_339_axes_0 = const()[name = string("input_339_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132622464)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132624576)))];
+            tensor<fp16, [1, 28, 1024]> input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = string("input_339_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132626688))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135772480))))[name = string("encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_6_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135772672)))];
+            tensor<fp16, [1, 28, 4096]> linear_55_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized, x = input_339_cast_fp16)[name = string("linear_55_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = string("input_343_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135780928))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138926720))))[name = string("encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138926912)))];
+            tensor<fp16, [1, 28, 1024]> linear_56_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized, x = input_343_cast_fp16)[name = string("linear_56_cast_fp16")];
+            fp16 var_1733_to_fp16 = const()[name = string("op_1733_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_1734_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1733_to_fp16)[name = string("op_1734_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1734_cast_fp16)[name = string("input_349_cast_fp16")];
+            tensor<int32, [1]> key_13_axes_0 = const()[name = string("key_13_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138929024)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138931136)))];
+            tensor<fp16, [1, 28, 1024]> key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = string("key_13_cast_fp16")];
+            bool input_351_interleave_0 = const()[name = string("input_351_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_351_cast_fp16 = concat(axis = var_68, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = string("input_351_cast_fp16")];
+            tensor<int32, [3]> var_1756_begin_0 = const()[name = string("op_1756_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_1756_end_0 = const()[name = string("op_1756_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1756_end_mask_0 = const()[name = string("op_1756_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_1756_cast_fp16 = slice_by_index(begin = var_1756_begin_0, end = var_1756_end_0, end_mask = var_1756_end_mask_0, x = cache_25_cast_fp16)[name = string("op_1756_cast_fp16")];
+            bool var_1762_interleave_0 = const()[name = string("op_1762_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1762_cast_fp16 = concat(axis = var_68, interleave = var_1762_interleave_0, values = (var_1756_cast_fp16, key_13_cast_fp16))[name = string("op_1762_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138933248))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139719744))))[name = string("encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139719936)))];
+            tensor<fp16, [1, 28, 1024]> linear_57_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized, x = key_13_cast_fp16)[name = string("linear_57_cast_fp16")];
+            tensor<int32, [4]> var_1767 = const()[name = string("op_1767"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_37_cast_fp16 = reshape(shape = var_1767, x = linear_57_cast_fp16)[name = string("q_37_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139722048))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140508544))))[name = string("encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140508736)))];
+            tensor<fp16, [1, 70, 1024]> linear_58_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_58_cast_fp16")];
+            tensor<int32, [4]> var_1772 = const()[name = string("op_1772"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_25_cast_fp16 = reshape(shape = var_1772, x = linear_58_cast_fp16)[name = string("k_25_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140510848))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141297344))))[name = string("encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141297536)))];
+            tensor<fp16, [1, 70, 1024]> linear_59_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_59_cast_fp16")];
+            tensor<int32, [4]> var_1777 = const()[name = string("op_1777"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_13_cast_fp16 = reshape(shape = var_1777, x = linear_59_cast_fp16)[name = string("v_13_cast_fp16")];
+            tensor<int32, [4]> value_21_perm_0 = const()[name = string("value_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141299648)))];
+            tensor<fp16, [1, 28, 8, 128]> var_1790_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = string("op_1790_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141301760)))];
+            tensor<fp16, [1, 28, 8, 128]> var_1792_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = string("op_1792_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_13_perm_0 = const()[name = string("q_with_bias_v_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_163_transpose_x_0 = const()[name = string("x_163_transpose_x_0"), val = bool(false)];
+            bool x_163_transpose_y_0 = const()[name = string("x_163_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_1794_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141303872))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141446272))))[name = string("op_1794_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1792_cast_fp16)[name = string("transpose_308")];
+            tensor<fp16, [1, 8, 28, 139]> x_163_cast_fp16 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = op_1794_to_fp16_quantized)[name = string("x_163_cast_fp16")];
+            tensor<int32, [8]> x_165_pad_0 = const()[name = string("x_165_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_165_mode_0 = const()[name = string("x_165_mode_0"), val = string("constant")];
+            fp16 const_157_to_fp16 = const()[name = string("const_157_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_165_cast_fp16 = pad(constant_val = const_157_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = string("x_165_cast_fp16")];
+            tensor<int32, [4]> var_1802 = const()[name = string("op_1802"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_167_cast_fp16 = reshape(shape = var_1802, x = x_165_cast_fp16)[name = string("x_167_cast_fp16")];
+            tensor<int32, [4]> var_1806_begin_0 = const()[name = string("op_1806_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1806_end_0 = const()[name = string("op_1806_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_1806_end_mask_0 = const()[name = string("op_1806_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_1806_cast_fp16 = slice_by_index(begin = var_1806_begin_0, end = var_1806_end_0, end_mask = var_1806_end_mask_0, x = x_167_cast_fp16)[name = string("op_1806_cast_fp16")];
+            tensor<int32, [4]> var_1807 = const()[name = string("op_1807"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_25_cast_fp16 = reshape(shape = var_1807, x = var_1806_cast_fp16)[name = string("matrix_bd_25_cast_fp16")];
+            bool matrix_ac_13_transpose_x_0 = const()[name = string("matrix_ac_13_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_13_transpose_y_0 = const()[name = string("matrix_ac_13_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_109 = transpose(perm = transpose_109_perm_0, x = k_25_cast_fp16)[name = string("transpose_306")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1790_cast_fp16)[name = string("transpose_307")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_108, y = transpose_109)[name = string("matrix_ac_13_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_27_begin_0 = const()[name = string("matrix_bd_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_27_end_0 = const()[name = string("matrix_bd_27_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_27_end_mask_0 = const()[name = string("matrix_bd_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25_cast_fp16)[name = string("matrix_bd_27_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_1816_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = string("op_1816_cast_fp16")];
+            fp16 _inversed_scores_25_y_0_to_fp16 = const()[name = string("_inversed_scores_25_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_25_cast_fp16 = mul(x = var_1816_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = string("_inversed_scores_25_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_27_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_11)[name = string("scores_27_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_1822_cast_fp16 = softmax(axis = var_59, x = scores_27_cast_fp16)[name = string("op_1822_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_353_cast_fp16 = select(a = var_44_to_fp16, b = var_1822_cast_fp16, cond = mask_11)[name = string("input_353_cast_fp16")];
+            bool x_169_transpose_x_0 = const()[name = string("x_169_transpose_x_0"), val = bool(false)];
+            bool x_169_transpose_y_0 = const()[name = string("x_169_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_13_cast_fp16)[name = string("transpose_305")];
+            tensor<fp16, [1, 8, 28, 128]> x_169_cast_fp16 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353_cast_fp16, y = value_21_cast_fp16)[name = string("x_169_cast_fp16")];
+            tensor<int32, [4]> var_1826_perm_0 = const()[name = string("op_1826_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1827 = const()[name = string("op_1827"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_1826_cast_fp16 = transpose(perm = var_1826_perm_0, x = x_169_cast_fp16)[name = string("transpose_304")];
+            tensor<fp16, [1, 28, 1024]> input_355_cast_fp16 = reshape(shape = var_1827, x = var_1826_cast_fp16)[name = string("input_355_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141446656))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142233152))))[name = string("encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142233344)))];
+            tensor<fp16, [1, 28, 1024]> linear_61_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized, x = input_355_cast_fp16)[name = string("linear_61_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = string("input_359_cast_fp16")];
+            tensor<int32, [1]> x_173_axes_0 = const()[name = string("x_173_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142235456)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142237568)))];
+            tensor<fp16, [1, 28, 1024]> x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = string("x_173_cast_fp16")];
+            tensor<int32, [3]> input_361_perm_0 = const()[name = string("input_361_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_363_pad_type_0 = const()[name = string("input_363_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_363_strides_0 = const()[name = string("input_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_363_pad_0 = const()[name = string("input_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_363_dilations_0 = const()[name = string("input_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_363_groups_0 = const()[name = string("input_363_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142239680))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144336896))))[name = string("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = string("transpose_303")];
+            tensor<fp16, [1, 2048, 28]> input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_361_cast_fp16)[name = string("input_363_cast_fp16")];
+            int32 x_175_split_num_splits_0 = const()[name = string("x_175_split_num_splits_0"), val = int32(2)];
+            int32 x_175_split_axis_0 = const()[name = string("x_175_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_175_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_175_split_cast_fp16_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363_cast_fp16)[name = string("x_175_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = string("x_175_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = string("x_175_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_365_cast_fp16 = select(a = var_44_to_fp16, b = x_175_cast_fp16, cond = var_575)[name = string("input_365_cast_fp16")];
+            bool new_x_27_interleave_0 = const()[name = string("new_x_27_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_27_cast_fp16 = concat(axis = var_59, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = string("new_x_27_cast_fp16")];
+            tensor<int32, [3]> var_1866_begin_0 = const()[name = string("op_1866_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_1866_end_0 = const()[name = string("op_1866_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_1866_end_mask_0 = const()[name = string("op_1866_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1866_cast_fp16 = slice_by_index(begin = var_1866_begin_0, end = var_1866_end_0, end_mask = var_1866_end_mask_0, x = new_x_27_cast_fp16)[name = string("op_1866_cast_fp16")];
+            string x_177_pad_type_0 = const()[name = string("x_177_pad_type_0"), val = string("valid")];
+            int32 x_177_groups_0 = const()[name = string("x_177_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_177_strides_0 = const()[name = string("x_177_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_177_pad_0 = const()[name = string("x_177_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_177_dilations_0 = const()[name = string("x_177_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144341056))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144350336))))[name = string("encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_177_cast_fp16 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_27_cast_fp16)[name = string("x_177_cast_fp16")];
+            tensor<int32, [3]> input_367_perm_0 = const()[name = string("input_367_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_179_axes_0 = const()[name = string("x_179_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144352448)))];
+            tensor<fp16, [1024]> encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144354560)))];
+            tensor<fp16, [1, 28, 1024]> input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = string("transpose_302")];
+            tensor<fp16, [1, 28, 1024]> x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = string("x_179_cast_fp16")];
+            tensor<int32, [3]> input_369_perm_0 = const()[name = string("input_369_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = string("transpose_301")];
+            tensor<fp16, [1, 1024, 28]> input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = string("input_371_cast_fp16")];
+            string x_181_pad_type_0 = const()[name = string("x_181_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_181_strides_0 = const()[name = string("x_181_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_181_pad_0 = const()[name = string("x_181_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_181_dilations_0 = const()[name = string("x_181_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_181_groups_0 = const()[name = string("x_181_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144356672))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145405312))))[name = string("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_181_cast_fp16 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_371_cast_fp16)[name = string("x_181_cast_fp16")];
+            tensor<int32, [3]> input_373_perm_0 = const()[name = string("input_373_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = string("transpose_300")];
+            tensor<fp16, [1, 28, 1024]> input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = string("input_375_cast_fp16")];
+            tensor<int32, [1]> input_377_axes_0 = const()[name = string("input_377_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145407424)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145409536)))];
+            tensor<fp16, [1, 28, 1024]> input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = string("input_377_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145411648))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148557440))))[name = string("encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_6_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148557632)))];
+            tensor<fp16, [1, 28, 4096]> linear_62_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized, x = input_377_cast_fp16)[name = string("linear_62_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = string("input_381_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148565888))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151711680))))[name = string("encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151711872)))];
+            tensor<fp16, [1, 28, 1024]> linear_63_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized, x = input_381_cast_fp16)[name = string("linear_63_cast_fp16")];
+            fp16 var_1909_to_fp16 = const()[name = string("op_1909_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_1910_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1909_to_fp16)[name = string("op_1910_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1910_cast_fp16)[name = string("input_387_cast_fp16")];
+            tensor<int32, [1]> input_389_axes_0 = const()[name = string("input_389_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151713984)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151716096)))];
+            tensor<fp16, [1, 28, 1024]> input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = string("input_389_cast_fp16")];
+            tensor<int32, [4]> cache_29_begin_0 = const()[name = string("cache_29_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_29_end_0 = const()[name = string("cache_29_end_0"), val = tensor<int32, [4]>([8, 1, 42, 1024])];
+            tensor<bool, [4]> cache_29_end_mask_0 = const()[name = string("cache_29_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_29_squeeze_mask_0 = const()[name = string("cache_29_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_29_cast_fp16 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_29_cast_fp16")];
+            tensor<int32, [4]> cache_31_begin_0 = const()[name = string("cache_31_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_31_end_0 = const()[name = string("cache_31_end_0"), val = tensor<int32, [4]>([8, 1, 1024, 8])];
+            tensor<bool, [4]> cache_31_end_mask_0 = const()[name = string("cache_31_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_31_squeeze_mask_0 = const()[name = string("cache_31_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_31_cast_fp16 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_31_cast_fp16")];
+            tensor<int32, [1]> input_391_axes_0 = const()[name = string("input_391_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151718208)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151720320)))];
+            tensor<fp16, [1, 28, 1024]> input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = string("input_391_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151722432))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154868224))))[name = string("encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_7_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154868416)))];
+            tensor<fp16, [1, 28, 4096]> linear_64_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized, x = input_391_cast_fp16)[name = string("linear_64_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = string("input_395_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154876672))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158022464))))[name = string("encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158022656)))];
+            tensor<fp16, [1, 28, 1024]> linear_65_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized, x = input_395_cast_fp16)[name = string("linear_65_cast_fp16")];
+            fp16 var_1946_to_fp16 = const()[name = string("op_1946_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_1947_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1946_to_fp16)[name = string("op_1947_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1947_cast_fp16)[name = string("input_401_cast_fp16")];
+            tensor<int32, [1]> key_15_axes_0 = const()[name = string("key_15_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158024768)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158026880)))];
+            tensor<fp16, [1, 28, 1024]> key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = string("key_15_cast_fp16")];
+            bool input_403_interleave_0 = const()[name = string("input_403_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_403_cast_fp16 = concat(axis = var_68, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = string("input_403_cast_fp16")];
+            tensor<int32, [3]> var_1969_begin_0 = const()[name = string("op_1969_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_1969_end_0 = const()[name = string("op_1969_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1969_end_mask_0 = const()[name = string("op_1969_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_1969_cast_fp16 = slice_by_index(begin = var_1969_begin_0, end = var_1969_end_0, end_mask = var_1969_end_mask_0, x = cache_29_cast_fp16)[name = string("op_1969_cast_fp16")];
+            bool var_1975_interleave_0 = const()[name = string("op_1975_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1975_cast_fp16 = concat(axis = var_68, interleave = var_1975_interleave_0, values = (var_1969_cast_fp16, key_15_cast_fp16))[name = string("op_1975_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158028992))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158815488))))[name = string("encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158815680)))];
+            tensor<fp16, [1, 28, 1024]> linear_66_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized, x = key_15_cast_fp16)[name = string("linear_66_cast_fp16")];
+            tensor<int32, [4]> var_1980 = const()[name = string("op_1980"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_43_cast_fp16 = reshape(shape = var_1980, x = linear_66_cast_fp16)[name = string("q_43_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158817792))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159604288))))[name = string("encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159604480)))];
+            tensor<fp16, [1, 70, 1024]> linear_67_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_67_cast_fp16")];
+            tensor<int32, [4]> var_1985 = const()[name = string("op_1985"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_29_cast_fp16 = reshape(shape = var_1985, x = linear_67_cast_fp16)[name = string("k_29_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159606592))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160393088))))[name = string("encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160393280)))];
+            tensor<fp16, [1, 70, 1024]> linear_68_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_68_cast_fp16")];
+            tensor<int32, [4]> var_1990 = const()[name = string("op_1990"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_15_cast_fp16 = reshape(shape = var_1990, x = linear_68_cast_fp16)[name = string("v_15_cast_fp16")];
+            tensor<int32, [4]> value_23_perm_0 = const()[name = string("value_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160395392)))];
+            tensor<fp16, [1, 28, 8, 128]> var_2003_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = string("op_2003_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160397504)))];
+            tensor<fp16, [1, 28, 8, 128]> var_2005_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = string("op_2005_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_15_perm_0 = const()[name = string("q_with_bias_v_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_189_transpose_x_0 = const()[name = string("x_189_transpose_x_0"), val = bool(false)];
+            bool x_189_transpose_y_0 = const()[name = string("x_189_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_2007_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160399616))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160542016))))[name = string("op_2007_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_2005_cast_fp16)[name = string("transpose_299")];
+            tensor<fp16, [1, 8, 28, 139]> x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = op_2007_to_fp16_quantized)[name = string("x_189_cast_fp16")];
+            tensor<int32, [8]> x_191_pad_0 = const()[name = string("x_191_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_191_mode_0 = const()[name = string("x_191_mode_0"), val = string("constant")];
+            fp16 const_170_to_fp16 = const()[name = string("const_170_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_191_cast_fp16 = pad(constant_val = const_170_to_fp16, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189_cast_fp16)[name = string("x_191_cast_fp16")];
+            tensor<int32, [4]> var_2015 = const()[name = string("op_2015"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_193_cast_fp16 = reshape(shape = var_2015, x = x_191_cast_fp16)[name = string("x_193_cast_fp16")];
+            tensor<int32, [4]> var_2019_begin_0 = const()[name = string("op_2019_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2019_end_0 = const()[name = string("op_2019_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_2019_end_mask_0 = const()[name = string("op_2019_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_2019_cast_fp16 = slice_by_index(begin = var_2019_begin_0, end = var_2019_end_0, end_mask = var_2019_end_mask_0, x = x_193_cast_fp16)[name = string("op_2019_cast_fp16")];
+            tensor<int32, [4]> var_2020 = const()[name = string("op_2020"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_29_cast_fp16 = reshape(shape = var_2020, x = var_2019_cast_fp16)[name = string("matrix_bd_29_cast_fp16")];
+            bool matrix_ac_15_transpose_x_0 = const()[name = string("matrix_ac_15_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_15_transpose_y_0 = const()[name = string("matrix_ac_15_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_111 = transpose(perm = transpose_111_perm_0, x = k_29_cast_fp16)[name = string("transpose_297")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_110 = transpose(perm = transpose_110_perm_0, x = var_2003_cast_fp16)[name = string("transpose_298")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_110, y = transpose_111)[name = string("matrix_ac_15_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_31_begin_0 = const()[name = string("matrix_bd_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_31_end_0 = const()[name = string("matrix_bd_31_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_31_end_mask_0 = const()[name = string("matrix_bd_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29_cast_fp16)[name = string("matrix_bd_31_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_2029_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = string("op_2029_cast_fp16")];
+            fp16 _inversed_scores_29_y_0_to_fp16 = const()[name = string("_inversed_scores_29_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_29_cast_fp16 = mul(x = var_2029_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = string("_inversed_scores_29_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_31_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_11)[name = string("scores_31_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_2035_cast_fp16 = softmax(axis = var_59, x = scores_31_cast_fp16)[name = string("op_2035_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_405_cast_fp16 = select(a = var_44_to_fp16, b = var_2035_cast_fp16, cond = mask_11)[name = string("input_405_cast_fp16")];
+            bool x_195_transpose_x_0 = const()[name = string("x_195_transpose_x_0"), val = bool(false)];
+            bool x_195_transpose_y_0 = const()[name = string("x_195_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_15_cast_fp16)[name = string("transpose_296")];
+            tensor<fp16, [1, 8, 28, 128]> x_195_cast_fp16 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405_cast_fp16, y = value_23_cast_fp16)[name = string("x_195_cast_fp16")];
+            tensor<int32, [4]> var_2039_perm_0 = const()[name = string("op_2039_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2040 = const()[name = string("op_2040"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_2039_cast_fp16 = transpose(perm = var_2039_perm_0, x = x_195_cast_fp16)[name = string("transpose_295")];
+            tensor<fp16, [1, 28, 1024]> input_407_cast_fp16 = reshape(shape = var_2040, x = var_2039_cast_fp16)[name = string("input_407_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160542400))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161328896))))[name = string("encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161329088)))];
+            tensor<fp16, [1, 28, 1024]> linear_70_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = string("linear_70_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = string("input_411_cast_fp16")];
+            tensor<int32, [1]> x_199_axes_0 = const()[name = string("x_199_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161331200)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161333312)))];
+            tensor<fp16, [1, 28, 1024]> x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = string("x_199_cast_fp16")];
+            tensor<int32, [3]> input_413_perm_0 = const()[name = string("input_413_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_415_pad_type_0 = const()[name = string("input_415_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_415_strides_0 = const()[name = string("input_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_415_pad_0 = const()[name = string("input_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_415_dilations_0 = const()[name = string("input_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_415_groups_0 = const()[name = string("input_415_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161335424))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163432640))))[name = string("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = string("transpose_294")];
+            tensor<fp16, [1, 2048, 28]> input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_413_cast_fp16)[name = string("input_415_cast_fp16")];
+            int32 x_201_split_num_splits_0 = const()[name = string("x_201_split_num_splits_0"), val = int32(2)];
+            int32 x_201_split_axis_0 = const()[name = string("x_201_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_201_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_201_split_cast_fp16_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415_cast_fp16)[name = string("x_201_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = string("x_201_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = string("x_201_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_417_cast_fp16 = select(a = var_44_to_fp16, b = x_201_cast_fp16, cond = var_575)[name = string("input_417_cast_fp16")];
+            bool new_x_31_interleave_0 = const()[name = string("new_x_31_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_31_cast_fp16 = concat(axis = var_59, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = string("new_x_31_cast_fp16")];
+            tensor<int32, [3]> var_2079_begin_0 = const()[name = string("op_2079_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_2079_end_0 = const()[name = string("op_2079_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_2079_end_mask_0 = const()[name = string("op_2079_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2079_cast_fp16 = slice_by_index(begin = var_2079_begin_0, end = var_2079_end_0, end_mask = var_2079_end_mask_0, x = new_x_31_cast_fp16)[name = string("op_2079_cast_fp16")];
+            string x_203_pad_type_0 = const()[name = string("x_203_pad_type_0"), val = string("valid")];
+            int32 x_203_groups_0 = const()[name = string("x_203_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_203_strides_0 = const()[name = string("x_203_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_203_pad_0 = const()[name = string("x_203_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_203_dilations_0 = const()[name = string("x_203_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163436800))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163446080))))[name = string("encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_31_cast_fp16)[name = string("x_203_cast_fp16")];
+            tensor<int32, [3]> input_419_perm_0 = const()[name = string("input_419_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_205_axes_0 = const()[name = string("x_205_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163448192)))];
+            tensor<fp16, [1024]> encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163450304)))];
+            tensor<fp16, [1, 28, 1024]> input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = string("transpose_293")];
+            tensor<fp16, [1, 28, 1024]> x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = string("x_205_cast_fp16")];
+            tensor<int32, [3]> input_421_perm_0 = const()[name = string("input_421_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = string("transpose_292")];
+            tensor<fp16, [1, 1024, 28]> input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = string("input_423_cast_fp16")];
+            string x_207_pad_type_0 = const()[name = string("x_207_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_207_strides_0 = const()[name = string("x_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_207_pad_0 = const()[name = string("x_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_207_dilations_0 = const()[name = string("x_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_207_groups_0 = const()[name = string("x_207_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163452416))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164501056))))[name = string("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_207_cast_fp16 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_423_cast_fp16)[name = string("x_207_cast_fp16")];
+            tensor<int32, [3]> input_425_perm_0 = const()[name = string("input_425_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = string("transpose_291")];
+            tensor<fp16, [1, 28, 1024]> input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = string("input_427_cast_fp16")];
+            tensor<int32, [1]> input_429_axes_0 = const()[name = string("input_429_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164503168)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164505280)))];
+            tensor<fp16, [1, 28, 1024]> input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = string("input_429_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164507392))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167653184))))[name = string("encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_7_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167653376)))];
+            tensor<fp16, [1, 28, 4096]> linear_71_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized, x = input_429_cast_fp16)[name = string("linear_71_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = string("input_433_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167661632))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170807424))))[name = string("encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170807616)))];
+            tensor<fp16, [1, 28, 1024]> linear_72_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized, x = input_433_cast_fp16)[name = string("linear_72_cast_fp16")];
+            fp16 var_2122_to_fp16 = const()[name = string("op_2122_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_2123_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_2122_to_fp16)[name = string("op_2123_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_2123_cast_fp16)[name = string("input_439_cast_fp16")];
+            tensor<int32, [1]> input_441_axes_0 = const()[name = string("input_441_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170809728)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170811840)))];
+            tensor<fp16, [1, 28, 1024]> input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = string("input_441_cast_fp16")];
+            tensor<int32, [4]> cache_33_begin_0 = const()[name = string("cache_33_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_33_end_0 = const()[name = string("cache_33_end_0"), val = tensor<int32, [4]>([9, 1, 42, 1024])];
+            tensor<bool, [4]> cache_33_end_mask_0 = const()[name = string("cache_33_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_33_squeeze_mask_0 = const()[name = string("cache_33_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_33_cast_fp16 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_33_cast_fp16")];
+            tensor<int32, [4]> cache_35_begin_0 = const()[name = string("cache_35_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_35_end_0 = const()[name = string("cache_35_end_0"), val = tensor<int32, [4]>([9, 1, 1024, 8])];
+            tensor<bool, [4]> cache_35_end_mask_0 = const()[name = string("cache_35_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_35_squeeze_mask_0 = const()[name = string("cache_35_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_35_cast_fp16 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_35_cast_fp16")];
+            tensor<int32, [1]> input_443_axes_0 = const()[name = string("input_443_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170813952)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170816064)))];
+            tensor<fp16, [1, 28, 1024]> input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = string("input_443_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170818176))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173963968))))[name = string("encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_8_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173964160)))];
+            tensor<fp16, [1, 28, 4096]> linear_73_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("linear_73_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = string("input_447_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173972416))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177118208))))[name = string("encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177118400)))];
+            tensor<fp16, [1, 28, 1024]> linear_74_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized, x = input_447_cast_fp16)[name = string("linear_74_cast_fp16")];
+            fp16 var_2159_to_fp16 = const()[name = string("op_2159_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_2160_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_2159_to_fp16)[name = string("op_2160_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_2160_cast_fp16)[name = string("input_453_cast_fp16")];
+            tensor<int32, [1]> key_17_axes_0 = const()[name = string("key_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177120512)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177122624)))];
+            tensor<fp16, [1, 28, 1024]> key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = string("key_17_cast_fp16")];
+            bool input_455_interleave_0 = const()[name = string("input_455_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_455_cast_fp16 = concat(axis = var_68, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = string("input_455_cast_fp16")];
+            tensor<int32, [3]> var_2182_begin_0 = const()[name = string("op_2182_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_2182_end_0 = const()[name = string("op_2182_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2182_end_mask_0 = const()[name = string("op_2182_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_2182_cast_fp16 = slice_by_index(begin = var_2182_begin_0, end = var_2182_end_0, end_mask = var_2182_end_mask_0, x = cache_33_cast_fp16)[name = string("op_2182_cast_fp16")];
+            bool var_2188_interleave_0 = const()[name = string("op_2188_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_2188_cast_fp16 = concat(axis = var_68, interleave = var_2188_interleave_0, values = (var_2182_cast_fp16, key_17_cast_fp16))[name = string("op_2188_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177124736))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177911232))))[name = string("encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177911424)))];
+            tensor<fp16, [1, 28, 1024]> linear_75_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized, x = key_17_cast_fp16)[name = string("linear_75_cast_fp16")];
+            tensor<int32, [4]> var_2193 = const()[name = string("op_2193"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_49_cast_fp16 = reshape(shape = var_2193, x = linear_75_cast_fp16)[name = string("q_49_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177913536))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178700032))))[name = string("encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178700224)))];
+            tensor<fp16, [1, 70, 1024]> linear_76_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_76_cast_fp16")];
+            tensor<int32, [4]> var_2198 = const()[name = string("op_2198"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_33_cast_fp16 = reshape(shape = var_2198, x = linear_76_cast_fp16)[name = string("k_33_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178702336))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179488832))))[name = string("encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179489024)))];
+            tensor<fp16, [1, 70, 1024]> linear_77_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_77_cast_fp16")];
+            tensor<int32, [4]> var_2203 = const()[name = string("op_2203"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_17_cast_fp16 = reshape(shape = var_2203, x = linear_77_cast_fp16)[name = string("v_17_cast_fp16")];
+            tensor<int32, [4]> value_25_perm_0 = const()[name = string("value_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179491136)))];
+            tensor<fp16, [1, 28, 8, 128]> var_2216_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = string("op_2216_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179493248)))];
+            tensor<fp16, [1, 28, 8, 128]> var_2218_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = string("op_2218_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_17_perm_0 = const()[name = string("q_with_bias_v_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_215_transpose_x_0 = const()[name = string("x_215_transpose_x_0"), val = bool(false)];
+            bool x_215_transpose_y_0 = const()[name = string("x_215_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_2220_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179495360))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179637760))))[name = string("op_2220_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2218_cast_fp16)[name = string("transpose_290")];
+            tensor<fp16, [1, 8, 28, 139]> x_215_cast_fp16 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = op_2220_to_fp16_quantized)[name = string("x_215_cast_fp16")];
+            tensor<int32, [8]> x_217_pad_0 = const()[name = string("x_217_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_217_mode_0 = const()[name = string("x_217_mode_0"), val = string("constant")];
+            fp16 const_183_to_fp16 = const()[name = string("const_183_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_217_cast_fp16 = pad(constant_val = const_183_to_fp16, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215_cast_fp16)[name = string("x_217_cast_fp16")];
+            tensor<int32, [4]> var_2228 = const()[name = string("op_2228"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_219_cast_fp16 = reshape(shape = var_2228, x = x_217_cast_fp16)[name = string("x_219_cast_fp16")];
+            tensor<int32, [4]> var_2232_begin_0 = const()[name = string("op_2232_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2232_end_0 = const()[name = string("op_2232_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_2232_end_mask_0 = const()[name = string("op_2232_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_2232_cast_fp16 = slice_by_index(begin = var_2232_begin_0, end = var_2232_end_0, end_mask = var_2232_end_mask_0, x = x_219_cast_fp16)[name = string("op_2232_cast_fp16")];
+            tensor<int32, [4]> var_2233 = const()[name = string("op_2233"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_33_cast_fp16 = reshape(shape = var_2233, x = var_2232_cast_fp16)[name = string("matrix_bd_33_cast_fp16")];
+            bool matrix_ac_17_transpose_x_0 = const()[name = string("matrix_ac_17_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_17_transpose_y_0 = const()[name = string("matrix_ac_17_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_113 = transpose(perm = transpose_113_perm_0, x = k_33_cast_fp16)[name = string("transpose_288")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_112 = transpose(perm = transpose_112_perm_0, x = var_2216_cast_fp16)[name = string("transpose_289")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_112, y = transpose_113)[name = string("matrix_ac_17_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_35_begin_0 = const()[name = string("matrix_bd_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_35_end_0 = const()[name = string("matrix_bd_35_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_35_end_mask_0 = const()[name = string("matrix_bd_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33_cast_fp16)[name = string("matrix_bd_35_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_2242_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = string("op_2242_cast_fp16")];
+            fp16 _inversed_scores_33_y_0_to_fp16 = const()[name = string("_inversed_scores_33_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_33_cast_fp16 = mul(x = var_2242_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = string("_inversed_scores_33_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_35_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_11)[name = string("scores_35_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_2248_cast_fp16 = softmax(axis = var_59, x = scores_35_cast_fp16)[name = string("op_2248_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_457_cast_fp16 = select(a = var_44_to_fp16, b = var_2248_cast_fp16, cond = mask_11)[name = string("input_457_cast_fp16")];
+            bool x_221_transpose_x_0 = const()[name = string("x_221_transpose_x_0"), val = bool(false)];
+            bool x_221_transpose_y_0 = const()[name = string("x_221_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_17_cast_fp16)[name = string("transpose_287")];
+            tensor<fp16, [1, 8, 28, 128]> x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457_cast_fp16, y = value_25_cast_fp16)[name = string("x_221_cast_fp16")];
+            tensor<int32, [4]> var_2252_perm_0 = const()[name = string("op_2252_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2253 = const()[name = string("op_2253"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_2252_cast_fp16 = transpose(perm = var_2252_perm_0, x = x_221_cast_fp16)[name = string("transpose_286")];
+            tensor<fp16, [1, 28, 1024]> input_459_cast_fp16 = reshape(shape = var_2253, x = var_2252_cast_fp16)[name = string("input_459_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179638144))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180424640))))[name = string("encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180424832)))];
+            tensor<fp16, [1, 28, 1024]> linear_79_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized, x = input_459_cast_fp16)[name = string("linear_79_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = string("input_463_cast_fp16")];
+            tensor<int32, [1]> x_225_axes_0 = const()[name = string("x_225_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180426944)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180429056)))];
+            tensor<fp16, [1, 28, 1024]> x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = string("x_225_cast_fp16")];
+            tensor<int32, [3]> input_465_perm_0 = const()[name = string("input_465_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_467_pad_type_0 = const()[name = string("input_467_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_467_strides_0 = const()[name = string("input_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_467_pad_0 = const()[name = string("input_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_467_dilations_0 = const()[name = string("input_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_467_groups_0 = const()[name = string("input_467_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180431168))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182528384))))[name = string("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = string("transpose_285")];
+            tensor<fp16, [1, 2048, 28]> input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_465_cast_fp16)[name = string("input_467_cast_fp16")];
+            int32 x_227_split_num_splits_0 = const()[name = string("x_227_split_num_splits_0"), val = int32(2)];
+            int32 x_227_split_axis_0 = const()[name = string("x_227_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_227_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_227_split_cast_fp16_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467_cast_fp16)[name = string("x_227_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = string("x_227_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = string("x_227_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_469_cast_fp16 = select(a = var_44_to_fp16, b = x_227_cast_fp16, cond = var_575)[name = string("input_469_cast_fp16")];
+            bool new_x_35_interleave_0 = const()[name = string("new_x_35_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_35_cast_fp16 = concat(axis = var_59, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = string("new_x_35_cast_fp16")];
+            tensor<int32, [3]> var_2292_begin_0 = const()[name = string("op_2292_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_2292_end_0 = const()[name = string("op_2292_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_2292_end_mask_0 = const()[name = string("op_2292_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2292_cast_fp16 = slice_by_index(begin = var_2292_begin_0, end = var_2292_end_0, end_mask = var_2292_end_mask_0, x = new_x_35_cast_fp16)[name = string("op_2292_cast_fp16")];
+            string x_229_pad_type_0 = const()[name = string("x_229_pad_type_0"), val = string("valid")];
+            int32 x_229_groups_0 = const()[name = string("x_229_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_229_strides_0 = const()[name = string("x_229_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_229_pad_0 = const()[name = string("x_229_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_229_dilations_0 = const()[name = string("x_229_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182532544))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182541824))))[name = string("encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_229_cast_fp16 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_35_cast_fp16)[name = string("x_229_cast_fp16")];
+            tensor<int32, [3]> input_471_perm_0 = const()[name = string("input_471_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_231_axes_0 = const()[name = string("x_231_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182543936)))];
+            tensor<fp16, [1024]> encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182546048)))];
+            tensor<fp16, [1, 28, 1024]> input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = string("transpose_284")];
+            tensor<fp16, [1, 28, 1024]> x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = string("x_231_cast_fp16")];
+            tensor<int32, [3]> input_473_perm_0 = const()[name = string("input_473_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = string("transpose_283")];
+            tensor<fp16, [1, 1024, 28]> input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = string("input_475_cast_fp16")];
+            string x_233_pad_type_0 = const()[name = string("x_233_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_233_strides_0 = const()[name = string("x_233_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_233_pad_0 = const()[name = string("x_233_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_233_dilations_0 = const()[name = string("x_233_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_233_groups_0 = const()[name = string("x_233_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182548160))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183596800))))[name = string("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_233_cast_fp16 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_475_cast_fp16)[name = string("x_233_cast_fp16")];
+            tensor<int32, [3]> input_477_perm_0 = const()[name = string("input_477_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = string("transpose_282")];
+            tensor<fp16, [1, 28, 1024]> input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = string("input_479_cast_fp16")];
+            tensor<int32, [1]> input_481_axes_0 = const()[name = string("input_481_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183598912)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183601024)))];
+            tensor<fp16, [1, 28, 1024]> input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = string("input_481_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183603136))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186748928))))[name = string("encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_8_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186749120)))];
+            tensor<fp16, [1, 28, 4096]> linear_80_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized, x = input_481_cast_fp16)[name = string("linear_80_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = string("input_485_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186757376))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189903168))))[name = string("encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189903360)))];
+            tensor<fp16, [1, 28, 1024]> linear_81_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized, x = input_485_cast_fp16)[name = string("linear_81_cast_fp16")];
+            fp16 var_2335_to_fp16 = const()[name = string("op_2335_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_2336_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2335_to_fp16)[name = string("op_2336_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2336_cast_fp16)[name = string("input_491_cast_fp16")];
+            tensor<int32, [1]> input_493_axes_0 = const()[name = string("input_493_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189905472)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189907584)))];
+            tensor<fp16, [1, 28, 1024]> input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = string("input_493_cast_fp16")];
+            tensor<int32, [4]> cache_37_begin_0 = const()[name = string("cache_37_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_37_end_0 = const()[name = string("cache_37_end_0"), val = tensor<int32, [4]>([10, 1, 42, 1024])];
+            tensor<bool, [4]> cache_37_end_mask_0 = const()[name = string("cache_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_37_squeeze_mask_0 = const()[name = string("cache_37_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_37_cast_fp16 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_37_cast_fp16")];
+            tensor<int32, [4]> cache_39_begin_0 = const()[name = string("cache_39_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_39_end_0 = const()[name = string("cache_39_end_0"), val = tensor<int32, [4]>([10, 1, 1024, 8])];
+            tensor<bool, [4]> cache_39_end_mask_0 = const()[name = string("cache_39_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_39_squeeze_mask_0 = const()[name = string("cache_39_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_39_cast_fp16 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_39_cast_fp16")];
+            tensor<int32, [1]> input_495_axes_0 = const()[name = string("input_495_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189909696)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189911808)))];
+            tensor<fp16, [1, 28, 1024]> input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = string("input_495_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189913920))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193059712))))[name = string("encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_9_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193059904)))];
+            tensor<fp16, [1, 28, 4096]> linear_82_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized, x = input_495_cast_fp16)[name = string("linear_82_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = string("input_499_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193068160))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196213952))))[name = string("encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196214144)))];
+            tensor<fp16, [1, 28, 1024]> linear_83_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized, x = input_499_cast_fp16)[name = string("linear_83_cast_fp16")];
+            fp16 var_2372_to_fp16 = const()[name = string("op_2372_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_2373_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2372_to_fp16)[name = string("op_2373_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2373_cast_fp16)[name = string("input_505_cast_fp16")];
+            tensor<int32, [1]> key_19_axes_0 = const()[name = string("key_19_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196216256)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196218368)))];
+            tensor<fp16, [1, 28, 1024]> key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = string("key_19_cast_fp16")];
+            bool input_507_interleave_0 = const()[name = string("input_507_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_507_cast_fp16 = concat(axis = var_68, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = string("input_507_cast_fp16")];
+            tensor<int32, [3]> var_2395_begin_0 = const()[name = string("op_2395_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_2395_end_0 = const()[name = string("op_2395_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2395_end_mask_0 = const()[name = string("op_2395_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_2395_cast_fp16 = slice_by_index(begin = var_2395_begin_0, end = var_2395_end_0, end_mask = var_2395_end_mask_0, x = cache_37_cast_fp16)[name = string("op_2395_cast_fp16")];
+            bool var_2401_interleave_0 = const()[name = string("op_2401_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_2401_cast_fp16 = concat(axis = var_68, interleave = var_2401_interleave_0, values = (var_2395_cast_fp16, key_19_cast_fp16))[name = string("op_2401_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196220480))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197006976))))[name = string("encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197007168)))];
+            tensor<fp16, [1, 28, 1024]> linear_84_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized, x = key_19_cast_fp16)[name = string("linear_84_cast_fp16")];
+            tensor<int32, [4]> var_2406 = const()[name = string("op_2406"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_55_cast_fp16 = reshape(shape = var_2406, x = linear_84_cast_fp16)[name = string("q_55_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197009280))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197795776))))[name = string("encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197795968)))];
+            tensor<fp16, [1, 70, 1024]> linear_85_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_85_cast_fp16")];
+            tensor<int32, [4]> var_2411 = const()[name = string("op_2411"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_37_cast_fp16 = reshape(shape = var_2411, x = linear_85_cast_fp16)[name = string("k_37_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197798080))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198584576))))[name = string("encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198584768)))];
+            tensor<fp16, [1, 70, 1024]> linear_86_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_86_cast_fp16")];
+            tensor<int32, [4]> var_2416 = const()[name = string("op_2416"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_19_cast_fp16 = reshape(shape = var_2416, x = linear_86_cast_fp16)[name = string("v_19_cast_fp16")];
+            tensor<int32, [4]> value_27_perm_0 = const()[name = string("value_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198586880)))];
+            tensor<fp16, [1, 28, 8, 128]> var_2429_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = string("op_2429_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198588992)))];
+            tensor<fp16, [1, 28, 8, 128]> var_2431_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = string("op_2431_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_19_perm_0 = const()[name = string("q_with_bias_v_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_241_transpose_x_0 = const()[name = string("x_241_transpose_x_0"), val = bool(false)];
+            bool x_241_transpose_y_0 = const()[name = string("x_241_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_2433_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198591104))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198733504))))[name = string("op_2433_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2431_cast_fp16)[name = string("transpose_281")];
+            tensor<fp16, [1, 8, 28, 139]> x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = op_2433_to_fp16_quantized)[name = string("x_241_cast_fp16")];
+            tensor<int32, [8]> x_243_pad_0 = const()[name = string("x_243_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_243_mode_0 = const()[name = string("x_243_mode_0"), val = string("constant")];
+            fp16 const_196_to_fp16 = const()[name = string("const_196_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_243_cast_fp16 = pad(constant_val = const_196_to_fp16, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241_cast_fp16)[name = string("x_243_cast_fp16")];
+            tensor<int32, [4]> var_2441 = const()[name = string("op_2441"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_245_cast_fp16 = reshape(shape = var_2441, x = x_243_cast_fp16)[name = string("x_245_cast_fp16")];
+            tensor<int32, [4]> var_2445_begin_0 = const()[name = string("op_2445_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2445_end_0 = const()[name = string("op_2445_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_2445_end_mask_0 = const()[name = string("op_2445_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_2445_cast_fp16 = slice_by_index(begin = var_2445_begin_0, end = var_2445_end_0, end_mask = var_2445_end_mask_0, x = x_245_cast_fp16)[name = string("op_2445_cast_fp16")];
+            tensor<int32, [4]> var_2446 = const()[name = string("op_2446"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_37_cast_fp16 = reshape(shape = var_2446, x = var_2445_cast_fp16)[name = string("matrix_bd_37_cast_fp16")];
+            bool matrix_ac_19_transpose_x_0 = const()[name = string("matrix_ac_19_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_19_transpose_y_0 = const()[name = string("matrix_ac_19_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_115 = transpose(perm = transpose_115_perm_0, x = k_37_cast_fp16)[name = string("transpose_279")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_114 = transpose(perm = transpose_114_perm_0, x = var_2429_cast_fp16)[name = string("transpose_280")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_19_cast_fp16 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_114, y = transpose_115)[name = string("matrix_ac_19_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_39_begin_0 = const()[name = string("matrix_bd_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_39_end_0 = const()[name = string("matrix_bd_39_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_39_end_mask_0 = const()[name = string("matrix_bd_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_39_cast_fp16 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37_cast_fp16)[name = string("matrix_bd_39_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_2455_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = string("op_2455_cast_fp16")];
+            fp16 _inversed_scores_37_y_0_to_fp16 = const()[name = string("_inversed_scores_37_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_37_cast_fp16 = mul(x = var_2455_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = string("_inversed_scores_37_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_39_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_11)[name = string("scores_39_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_2461_cast_fp16 = softmax(axis = var_59, x = scores_39_cast_fp16)[name = string("op_2461_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_509_cast_fp16 = select(a = var_44_to_fp16, b = var_2461_cast_fp16, cond = mask_11)[name = string("input_509_cast_fp16")];
+            bool x_247_transpose_x_0 = const()[name = string("x_247_transpose_x_0"), val = bool(false)];
+            bool x_247_transpose_y_0 = const()[name = string("x_247_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_19_cast_fp16)[name = string("transpose_278")];
+            tensor<fp16, [1, 8, 28, 128]> x_247_cast_fp16 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509_cast_fp16, y = value_27_cast_fp16)[name = string("x_247_cast_fp16")];
+            tensor<int32, [4]> var_2465_perm_0 = const()[name = string("op_2465_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2466 = const()[name = string("op_2466"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_2465_cast_fp16 = transpose(perm = var_2465_perm_0, x = x_247_cast_fp16)[name = string("transpose_277")];
+            tensor<fp16, [1, 28, 1024]> input_511_cast_fp16 = reshape(shape = var_2466, x = var_2465_cast_fp16)[name = string("input_511_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198733888))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199520384))))[name = string("encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199520576)))];
+            tensor<fp16, [1, 28, 1024]> linear_88_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized, x = input_511_cast_fp16)[name = string("linear_88_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = string("input_515_cast_fp16")];
+            tensor<int32, [1]> x_251_axes_0 = const()[name = string("x_251_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199522688)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199524800)))];
+            tensor<fp16, [1, 28, 1024]> x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = string("x_251_cast_fp16")];
+            tensor<int32, [3]> input_517_perm_0 = const()[name = string("input_517_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_519_pad_type_0 = const()[name = string("input_519_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_519_strides_0 = const()[name = string("input_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_519_pad_0 = const()[name = string("input_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_519_dilations_0 = const()[name = string("input_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_519_groups_0 = const()[name = string("input_519_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199526912))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201624128))))[name = string("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = string("transpose_276")];
+            tensor<fp16, [1, 2048, 28]> input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_517_cast_fp16)[name = string("input_519_cast_fp16")];
+            int32 x_253_split_num_splits_0 = const()[name = string("x_253_split_num_splits_0"), val = int32(2)];
+            int32 x_253_split_axis_0 = const()[name = string("x_253_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_253_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_253_split_cast_fp16_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519_cast_fp16)[name = string("x_253_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = string("x_253_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = string("x_253_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_521_cast_fp16 = select(a = var_44_to_fp16, b = x_253_cast_fp16, cond = var_575)[name = string("input_521_cast_fp16")];
+            bool new_x_39_interleave_0 = const()[name = string("new_x_39_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_39_cast_fp16 = concat(axis = var_59, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = string("new_x_39_cast_fp16")];
+            tensor<int32, [3]> var_2505_begin_0 = const()[name = string("op_2505_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_2505_end_0 = const()[name = string("op_2505_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_2505_end_mask_0 = const()[name = string("op_2505_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2505_cast_fp16 = slice_by_index(begin = var_2505_begin_0, end = var_2505_end_0, end_mask = var_2505_end_mask_0, x = new_x_39_cast_fp16)[name = string("op_2505_cast_fp16")];
+            string x_255_pad_type_0 = const()[name = string("x_255_pad_type_0"), val = string("valid")];
+            int32 x_255_groups_0 = const()[name = string("x_255_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_255_strides_0 = const()[name = string("x_255_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_255_pad_0 = const()[name = string("x_255_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_255_dilations_0 = const()[name = string("x_255_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201628288))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201637568))))[name = string("encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_39_cast_fp16)[name = string("x_255_cast_fp16")];
+            tensor<int32, [3]> input_523_perm_0 = const()[name = string("input_523_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_257_axes_0 = const()[name = string("x_257_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201639680)))];
+            tensor<fp16, [1024]> encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201641792)))];
+            tensor<fp16, [1, 28, 1024]> input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = string("transpose_275")];
+            tensor<fp16, [1, 28, 1024]> x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = string("x_257_cast_fp16")];
+            tensor<int32, [3]> input_525_perm_0 = const()[name = string("input_525_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = string("transpose_274")];
+            tensor<fp16, [1, 1024, 28]> input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = string("input_527_cast_fp16")];
+            string x_259_pad_type_0 = const()[name = string("x_259_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_259_strides_0 = const()[name = string("x_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_259_pad_0 = const()[name = string("x_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_259_dilations_0 = const()[name = string("x_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_259_groups_0 = const()[name = string("x_259_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201643904))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202692544))))[name = string("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_527_cast_fp16)[name = string("x_259_cast_fp16")];
+            tensor<int32, [3]> input_529_perm_0 = const()[name = string("input_529_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = string("transpose_273")];
+            tensor<fp16, [1, 28, 1024]> input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = string("input_531_cast_fp16")];
+            tensor<int32, [1]> input_533_axes_0 = const()[name = string("input_533_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202694656)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202696768)))];
+            tensor<fp16, [1, 28, 1024]> input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = string("input_533_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202698880))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205844672))))[name = string("encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_9_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205844864)))];
+            tensor<fp16, [1, 28, 4096]> linear_89_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized, x = input_533_cast_fp16)[name = string("linear_89_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = string("input_537_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205853120))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208998912))))[name = string("encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208999104)))];
+            tensor<fp16, [1, 28, 1024]> linear_90_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized, x = input_537_cast_fp16)[name = string("linear_90_cast_fp16")];
+            fp16 var_2548_to_fp16 = const()[name = string("op_2548_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_2549_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2548_to_fp16)[name = string("op_2549_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2549_cast_fp16)[name = string("input_543_cast_fp16")];
+            tensor<int32, [1]> input_545_axes_0 = const()[name = string("input_545_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209001216)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209003328)))];
+            tensor<fp16, [1, 28, 1024]> input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = string("input_545_cast_fp16")];
+            tensor<int32, [4]> cache_41_begin_0 = const()[name = string("cache_41_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_41_end_0 = const()[name = string("cache_41_end_0"), val = tensor<int32, [4]>([11, 1, 42, 1024])];
+            tensor<bool, [4]> cache_41_end_mask_0 = const()[name = string("cache_41_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_41_squeeze_mask_0 = const()[name = string("cache_41_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_41_cast_fp16 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_41_cast_fp16")];
+            tensor<int32, [4]> cache_43_begin_0 = const()[name = string("cache_43_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_43_end_0 = const()[name = string("cache_43_end_0"), val = tensor<int32, [4]>([11, 1, 1024, 8])];
+            tensor<bool, [4]> cache_43_end_mask_0 = const()[name = string("cache_43_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_43_squeeze_mask_0 = const()[name = string("cache_43_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_43_cast_fp16 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_43_cast_fp16")];
+            tensor<int32, [1]> input_547_axes_0 = const()[name = string("input_547_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209005440)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209007552)))];
+            tensor<fp16, [1, 28, 1024]> input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = string("input_547_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209009664))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212155456))))[name = string("encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_10_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212155648)))];
+            tensor<fp16, [1, 28, 4096]> linear_91_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized, x = input_547_cast_fp16)[name = string("linear_91_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = string("input_551_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212163904))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215309696))))[name = string("encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215309888)))];
+            tensor<fp16, [1, 28, 1024]> linear_92_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized, x = input_551_cast_fp16)[name = string("linear_92_cast_fp16")];
+            fp16 var_2585_to_fp16 = const()[name = string("op_2585_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_2586_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2585_to_fp16)[name = string("op_2586_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2586_cast_fp16)[name = string("input_557_cast_fp16")];
+            tensor<int32, [1]> key_21_axes_0 = const()[name = string("key_21_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215312000)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215314112)))];
+            tensor<fp16, [1, 28, 1024]> key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = string("key_21_cast_fp16")];
+            bool input_559_interleave_0 = const()[name = string("input_559_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_559_cast_fp16 = concat(axis = var_68, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = string("input_559_cast_fp16")];
+            tensor<int32, [3]> var_2608_begin_0 = const()[name = string("op_2608_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_2608_end_0 = const()[name = string("op_2608_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2608_end_mask_0 = const()[name = string("op_2608_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_2608_cast_fp16 = slice_by_index(begin = var_2608_begin_0, end = var_2608_end_0, end_mask = var_2608_end_mask_0, x = cache_41_cast_fp16)[name = string("op_2608_cast_fp16")];
+            bool var_2614_interleave_0 = const()[name = string("op_2614_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_2614_cast_fp16 = concat(axis = var_68, interleave = var_2614_interleave_0, values = (var_2608_cast_fp16, key_21_cast_fp16))[name = string("op_2614_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215316224))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216102720))))[name = string("encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216102912)))];
+            tensor<fp16, [1, 28, 1024]> linear_93_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized, x = key_21_cast_fp16)[name = string("linear_93_cast_fp16")];
+            tensor<int32, [4]> var_2619 = const()[name = string("op_2619"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_61_cast_fp16 = reshape(shape = var_2619, x = linear_93_cast_fp16)[name = string("q_61_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216105024))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216891520))))[name = string("encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216891712)))];
+            tensor<fp16, [1, 70, 1024]> linear_94_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_94_cast_fp16")];
+            tensor<int32, [4]> var_2624 = const()[name = string("op_2624"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_41_cast_fp16 = reshape(shape = var_2624, x = linear_94_cast_fp16)[name = string("k_41_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216893824))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217680320))))[name = string("encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217680512)))];
+            tensor<fp16, [1, 70, 1024]> linear_95_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_95_cast_fp16")];
+            tensor<int32, [4]> var_2629 = const()[name = string("op_2629"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_21_cast_fp16 = reshape(shape = var_2629, x = linear_95_cast_fp16)[name = string("v_21_cast_fp16")];
+            tensor<int32, [4]> value_29_perm_0 = const()[name = string("value_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217682624)))];
+            tensor<fp16, [1, 28, 8, 128]> var_2642_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = string("op_2642_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217684736)))];
+            tensor<fp16, [1, 28, 8, 128]> var_2644_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = string("op_2644_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_21_perm_0 = const()[name = string("q_with_bias_v_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_267_transpose_x_0 = const()[name = string("x_267_transpose_x_0"), val = bool(false)];
+            bool x_267_transpose_y_0 = const()[name = string("x_267_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_2646_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217686848))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217829248))))[name = string("op_2646_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2644_cast_fp16)[name = string("transpose_272")];
+            tensor<fp16, [1, 8, 28, 139]> x_267_cast_fp16 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = op_2646_to_fp16_quantized)[name = string("x_267_cast_fp16")];
+            tensor<int32, [8]> x_269_pad_0 = const()[name = string("x_269_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_269_mode_0 = const()[name = string("x_269_mode_0"), val = string("constant")];
+            fp16 const_209_to_fp16 = const()[name = string("const_209_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_269_cast_fp16 = pad(constant_val = const_209_to_fp16, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267_cast_fp16)[name = string("x_269_cast_fp16")];
+            tensor<int32, [4]> var_2654 = const()[name = string("op_2654"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_271_cast_fp16 = reshape(shape = var_2654, x = x_269_cast_fp16)[name = string("x_271_cast_fp16")];
+            tensor<int32, [4]> var_2658_begin_0 = const()[name = string("op_2658_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2658_end_0 = const()[name = string("op_2658_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_2658_end_mask_0 = const()[name = string("op_2658_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_2658_cast_fp16 = slice_by_index(begin = var_2658_begin_0, end = var_2658_end_0, end_mask = var_2658_end_mask_0, x = x_271_cast_fp16)[name = string("op_2658_cast_fp16")];
+            tensor<int32, [4]> var_2659 = const()[name = string("op_2659"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_41_cast_fp16 = reshape(shape = var_2659, x = var_2658_cast_fp16)[name = string("matrix_bd_41_cast_fp16")];
+            bool matrix_ac_21_transpose_x_0 = const()[name = string("matrix_ac_21_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_21_transpose_y_0 = const()[name = string("matrix_ac_21_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_117 = transpose(perm = transpose_117_perm_0, x = k_41_cast_fp16)[name = string("transpose_270")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_116 = transpose(perm = transpose_116_perm_0, x = var_2642_cast_fp16)[name = string("transpose_271")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_21_cast_fp16 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_116, y = transpose_117)[name = string("matrix_ac_21_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_43_begin_0 = const()[name = string("matrix_bd_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_43_end_0 = const()[name = string("matrix_bd_43_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_43_end_mask_0 = const()[name = string("matrix_bd_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_43_cast_fp16 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41_cast_fp16)[name = string("matrix_bd_43_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_2668_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = string("op_2668_cast_fp16")];
+            fp16 _inversed_scores_41_y_0_to_fp16 = const()[name = string("_inversed_scores_41_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_41_cast_fp16 = mul(x = var_2668_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = string("_inversed_scores_41_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_43_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_11)[name = string("scores_43_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_2674_cast_fp16 = softmax(axis = var_59, x = scores_43_cast_fp16)[name = string("op_2674_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_561_cast_fp16 = select(a = var_44_to_fp16, b = var_2674_cast_fp16, cond = mask_11)[name = string("input_561_cast_fp16")];
+            bool x_273_transpose_x_0 = const()[name = string("x_273_transpose_x_0"), val = bool(false)];
+            bool x_273_transpose_y_0 = const()[name = string("x_273_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_21_cast_fp16)[name = string("transpose_269")];
+            tensor<fp16, [1, 8, 28, 128]> x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561_cast_fp16, y = value_29_cast_fp16)[name = string("x_273_cast_fp16")];
+            tensor<int32, [4]> var_2678_perm_0 = const()[name = string("op_2678_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2679 = const()[name = string("op_2679"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_2678_cast_fp16 = transpose(perm = var_2678_perm_0, x = x_273_cast_fp16)[name = string("transpose_268")];
+            tensor<fp16, [1, 28, 1024]> input_563_cast_fp16 = reshape(shape = var_2679, x = var_2678_cast_fp16)[name = string("input_563_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217829632))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218616128))))[name = string("encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218616320)))];
+            tensor<fp16, [1, 28, 1024]> linear_97_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized, x = input_563_cast_fp16)[name = string("linear_97_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = string("input_567_cast_fp16")];
+            tensor<int32, [1]> x_277_axes_0 = const()[name = string("x_277_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218618432)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218620544)))];
+            tensor<fp16, [1, 28, 1024]> x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = string("x_277_cast_fp16")];
+            tensor<int32, [3]> input_569_perm_0 = const()[name = string("input_569_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_571_pad_type_0 = const()[name = string("input_571_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_571_strides_0 = const()[name = string("input_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_571_pad_0 = const()[name = string("input_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_571_dilations_0 = const()[name = string("input_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_571_groups_0 = const()[name = string("input_571_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218622656))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220719872))))[name = string("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = string("transpose_267")];
+            tensor<fp16, [1, 2048, 28]> input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_569_cast_fp16)[name = string("input_571_cast_fp16")];
+            int32 x_279_split_num_splits_0 = const()[name = string("x_279_split_num_splits_0"), val = int32(2)];
+            int32 x_279_split_axis_0 = const()[name = string("x_279_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_279_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_279_split_cast_fp16_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571_cast_fp16)[name = string("x_279_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = string("x_279_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = string("x_279_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_573_cast_fp16 = select(a = var_44_to_fp16, b = x_279_cast_fp16, cond = var_575)[name = string("input_573_cast_fp16")];
+            bool new_x_43_interleave_0 = const()[name = string("new_x_43_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_43_cast_fp16 = concat(axis = var_59, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = string("new_x_43_cast_fp16")];
+            tensor<int32, [3]> var_2718_begin_0 = const()[name = string("op_2718_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_2718_end_0 = const()[name = string("op_2718_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_2718_end_mask_0 = const()[name = string("op_2718_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2718_cast_fp16 = slice_by_index(begin = var_2718_begin_0, end = var_2718_end_0, end_mask = var_2718_end_mask_0, x = new_x_43_cast_fp16)[name = string("op_2718_cast_fp16")];
+            string x_281_pad_type_0 = const()[name = string("x_281_pad_type_0"), val = string("valid")];
+            int32 x_281_groups_0 = const()[name = string("x_281_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_281_strides_0 = const()[name = string("x_281_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_281_pad_0 = const()[name = string("x_281_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_281_dilations_0 = const()[name = string("x_281_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220724032))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220733312))))[name = string("encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_281_cast_fp16 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_43_cast_fp16)[name = string("x_281_cast_fp16")];
+            tensor<int32, [3]> input_575_perm_0 = const()[name = string("input_575_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_283_axes_0 = const()[name = string("x_283_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220735424)))];
+            tensor<fp16, [1024]> encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220737536)))];
+            tensor<fp16, [1, 28, 1024]> input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = string("transpose_266")];
+            tensor<fp16, [1, 28, 1024]> x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = string("x_283_cast_fp16")];
+            tensor<int32, [3]> input_577_perm_0 = const()[name = string("input_577_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = string("transpose_265")];
+            tensor<fp16, [1, 1024, 28]> input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = string("input_579_cast_fp16")];
+            string x_285_pad_type_0 = const()[name = string("x_285_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_285_strides_0 = const()[name = string("x_285_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_285_pad_0 = const()[name = string("x_285_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_285_dilations_0 = const()[name = string("x_285_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_285_groups_0 = const()[name = string("x_285_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220739648))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221788288))))[name = string("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_579_cast_fp16)[name = string("x_285_cast_fp16")];
+            tensor<int32, [3]> input_581_perm_0 = const()[name = string("input_581_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = string("transpose_264")];
+            tensor<fp16, [1, 28, 1024]> input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = string("input_583_cast_fp16")];
+            tensor<int32, [1]> input_585_axes_0 = const()[name = string("input_585_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221790400)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221792512)))];
+            tensor<fp16, [1, 28, 1024]> input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = string("input_585_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221794624))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224940416))))[name = string("encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_10_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224940608)))];
+            tensor<fp16, [1, 28, 4096]> linear_98_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized, x = input_585_cast_fp16)[name = string("linear_98_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = string("input_589_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224948864))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228094656))))[name = string("encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228094848)))];
+            tensor<fp16, [1, 28, 1024]> linear_99_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized, x = input_589_cast_fp16)[name = string("linear_99_cast_fp16")];
+            fp16 var_2761_to_fp16 = const()[name = string("op_2761_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_2762_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2761_to_fp16)[name = string("op_2762_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2762_cast_fp16)[name = string("input_595_cast_fp16")];
+            tensor<int32, [1]> input_597_axes_0 = const()[name = string("input_597_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228096960)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228099072)))];
+            tensor<fp16, [1, 28, 1024]> input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = string("input_597_cast_fp16")];
+            tensor<int32, [4]> cache_45_begin_0 = const()[name = string("cache_45_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_45_end_0 = const()[name = string("cache_45_end_0"), val = tensor<int32, [4]>([12, 1, 42, 1024])];
+            tensor<bool, [4]> cache_45_end_mask_0 = const()[name = string("cache_45_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_45_squeeze_mask_0 = const()[name = string("cache_45_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_45_cast_fp16 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_45_cast_fp16")];
+            tensor<int32, [4]> cache_47_begin_0 = const()[name = string("cache_47_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_47_end_0 = const()[name = string("cache_47_end_0"), val = tensor<int32, [4]>([12, 1, 1024, 8])];
+            tensor<bool, [4]> cache_47_end_mask_0 = const()[name = string("cache_47_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_47_squeeze_mask_0 = const()[name = string("cache_47_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_47_cast_fp16 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_47_cast_fp16")];
+            tensor<int32, [1]> input_599_axes_0 = const()[name = string("input_599_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228101184)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228103296)))];
+            tensor<fp16, [1, 28, 1024]> input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = string("input_599_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228105408))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231251200))))[name = string("encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_11_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231251392)))];
+            tensor<fp16, [1, 28, 4096]> linear_100_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized, x = input_599_cast_fp16)[name = string("linear_100_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = string("input_603_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231259648))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234405440))))[name = string("encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234405632)))];
+            tensor<fp16, [1, 28, 1024]> linear_101_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized, x = input_603_cast_fp16)[name = string("linear_101_cast_fp16")];
+            fp16 var_2798_to_fp16 = const()[name = string("op_2798_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_2799_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2798_to_fp16)[name = string("op_2799_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2799_cast_fp16)[name = string("input_609_cast_fp16")];
+            tensor<int32, [1]> key_23_axes_0 = const()[name = string("key_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234407744)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234409856)))];
+            tensor<fp16, [1, 28, 1024]> key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = string("key_23_cast_fp16")];
+            bool input_611_interleave_0 = const()[name = string("input_611_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_611_cast_fp16 = concat(axis = var_68, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = string("input_611_cast_fp16")];
+            tensor<int32, [3]> var_2821_begin_0 = const()[name = string("op_2821_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_2821_end_0 = const()[name = string("op_2821_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2821_end_mask_0 = const()[name = string("op_2821_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_2821_cast_fp16 = slice_by_index(begin = var_2821_begin_0, end = var_2821_end_0, end_mask = var_2821_end_mask_0, x = cache_45_cast_fp16)[name = string("op_2821_cast_fp16")];
+            bool var_2827_interleave_0 = const()[name = string("op_2827_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_2827_cast_fp16 = concat(axis = var_68, interleave = var_2827_interleave_0, values = (var_2821_cast_fp16, key_23_cast_fp16))[name = string("op_2827_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234411968))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235198464))))[name = string("encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235198656)))];
+            tensor<fp16, [1, 28, 1024]> linear_102_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized, x = key_23_cast_fp16)[name = string("linear_102_cast_fp16")];
+            tensor<int32, [4]> var_2832 = const()[name = string("op_2832"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_67_cast_fp16 = reshape(shape = var_2832, x = linear_102_cast_fp16)[name = string("q_67_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235200768))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235987264))))[name = string("encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235987456)))];
+            tensor<fp16, [1, 70, 1024]> linear_103_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_103_cast_fp16")];
+            tensor<int32, [4]> var_2837 = const()[name = string("op_2837"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_45_cast_fp16 = reshape(shape = var_2837, x = linear_103_cast_fp16)[name = string("k_45_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235989568))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236776064))))[name = string("encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236776256)))];
+            tensor<fp16, [1, 70, 1024]> linear_104_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_104_cast_fp16")];
+            tensor<int32, [4]> var_2842 = const()[name = string("op_2842"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_23_cast_fp16 = reshape(shape = var_2842, x = linear_104_cast_fp16)[name = string("v_23_cast_fp16")];
+            tensor<int32, [4]> value_31_perm_0 = const()[name = string("value_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236778368)))];
+            tensor<fp16, [1, 28, 8, 128]> var_2855_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = string("op_2855_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236780480)))];
+            tensor<fp16, [1, 28, 8, 128]> var_2857_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = string("op_2857_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_23_perm_0 = const()[name = string("q_with_bias_v_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_293_transpose_x_0 = const()[name = string("x_293_transpose_x_0"), val = bool(false)];
+            bool x_293_transpose_y_0 = const()[name = string("x_293_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_2859_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236782592))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236924992))))[name = string("op_2859_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2857_cast_fp16)[name = string("transpose_263")];
+            tensor<fp16, [1, 8, 28, 139]> x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = op_2859_to_fp16_quantized)[name = string("x_293_cast_fp16")];
+            tensor<int32, [8]> x_295_pad_0 = const()[name = string("x_295_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_295_mode_0 = const()[name = string("x_295_mode_0"), val = string("constant")];
+            fp16 const_222_to_fp16 = const()[name = string("const_222_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_295_cast_fp16 = pad(constant_val = const_222_to_fp16, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293_cast_fp16)[name = string("x_295_cast_fp16")];
+            tensor<int32, [4]> var_2867 = const()[name = string("op_2867"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_297_cast_fp16 = reshape(shape = var_2867, x = x_295_cast_fp16)[name = string("x_297_cast_fp16")];
+            tensor<int32, [4]> var_2871_begin_0 = const()[name = string("op_2871_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2871_end_0 = const()[name = string("op_2871_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_2871_end_mask_0 = const()[name = string("op_2871_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_2871_cast_fp16 = slice_by_index(begin = var_2871_begin_0, end = var_2871_end_0, end_mask = var_2871_end_mask_0, x = x_297_cast_fp16)[name = string("op_2871_cast_fp16")];
+            tensor<int32, [4]> var_2872 = const()[name = string("op_2872"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_45_cast_fp16 = reshape(shape = var_2872, x = var_2871_cast_fp16)[name = string("matrix_bd_45_cast_fp16")];
+            bool matrix_ac_23_transpose_x_0 = const()[name = string("matrix_ac_23_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_23_transpose_y_0 = const()[name = string("matrix_ac_23_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_119 = transpose(perm = transpose_119_perm_0, x = k_45_cast_fp16)[name = string("transpose_261")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_118 = transpose(perm = transpose_118_perm_0, x = var_2855_cast_fp16)[name = string("transpose_262")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_23_cast_fp16 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_118, y = transpose_119)[name = string("matrix_ac_23_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_47_begin_0 = const()[name = string("matrix_bd_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_47_end_0 = const()[name = string("matrix_bd_47_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_47_end_mask_0 = const()[name = string("matrix_bd_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_47_cast_fp16 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45_cast_fp16)[name = string("matrix_bd_47_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_2881_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = string("op_2881_cast_fp16")];
+            fp16 _inversed_scores_45_y_0_to_fp16 = const()[name = string("_inversed_scores_45_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_45_cast_fp16 = mul(x = var_2881_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = string("_inversed_scores_45_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_47_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_11)[name = string("scores_47_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_2887_cast_fp16 = softmax(axis = var_59, x = scores_47_cast_fp16)[name = string("op_2887_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_613_cast_fp16 = select(a = var_44_to_fp16, b = var_2887_cast_fp16, cond = mask_11)[name = string("input_613_cast_fp16")];
+            bool x_299_transpose_x_0 = const()[name = string("x_299_transpose_x_0"), val = bool(false)];
+            bool x_299_transpose_y_0 = const()[name = string("x_299_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_23_cast_fp16)[name = string("transpose_260")];
+            tensor<fp16, [1, 8, 28, 128]> x_299_cast_fp16 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613_cast_fp16, y = value_31_cast_fp16)[name = string("x_299_cast_fp16")];
+            tensor<int32, [4]> var_2891_perm_0 = const()[name = string("op_2891_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2892 = const()[name = string("op_2892"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_2891_cast_fp16 = transpose(perm = var_2891_perm_0, x = x_299_cast_fp16)[name = string("transpose_259")];
+            tensor<fp16, [1, 28, 1024]> input_615_cast_fp16 = reshape(shape = var_2892, x = var_2891_cast_fp16)[name = string("input_615_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236925376))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237711872))))[name = string("encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237712064)))];
+            tensor<fp16, [1, 28, 1024]> linear_106_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized, x = input_615_cast_fp16)[name = string("linear_106_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = string("input_619_cast_fp16")];
+            tensor<int32, [1]> x_303_axes_0 = const()[name = string("x_303_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237714176)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237716288)))];
+            tensor<fp16, [1, 28, 1024]> x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = string("x_303_cast_fp16")];
+            tensor<int32, [3]> input_621_perm_0 = const()[name = string("input_621_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_623_pad_type_0 = const()[name = string("input_623_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_623_strides_0 = const()[name = string("input_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_623_pad_0 = const()[name = string("input_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_623_dilations_0 = const()[name = string("input_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_623_groups_0 = const()[name = string("input_623_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237718400))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239815616))))[name = string("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = string("transpose_258")];
+            tensor<fp16, [1, 2048, 28]> input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_621_cast_fp16)[name = string("input_623_cast_fp16")];
+            int32 x_305_split_num_splits_0 = const()[name = string("x_305_split_num_splits_0"), val = int32(2)];
+            int32 x_305_split_axis_0 = const()[name = string("x_305_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_305_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623_cast_fp16)[name = string("x_305_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = string("x_305_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = string("x_305_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_625_cast_fp16 = select(a = var_44_to_fp16, b = x_305_cast_fp16, cond = var_575)[name = string("input_625_cast_fp16")];
+            bool new_x_47_interleave_0 = const()[name = string("new_x_47_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_47_cast_fp16 = concat(axis = var_59, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = string("new_x_47_cast_fp16")];
+            tensor<int32, [3]> var_2931_begin_0 = const()[name = string("op_2931_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_2931_end_0 = const()[name = string("op_2931_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_2931_end_mask_0 = const()[name = string("op_2931_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2931_cast_fp16 = slice_by_index(begin = var_2931_begin_0, end = var_2931_end_0, end_mask = var_2931_end_mask_0, x = new_x_47_cast_fp16)[name = string("op_2931_cast_fp16")];
+            string x_307_pad_type_0 = const()[name = string("x_307_pad_type_0"), val = string("valid")];
+            int32 x_307_groups_0 = const()[name = string("x_307_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_307_strides_0 = const()[name = string("x_307_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_307_pad_0 = const()[name = string("x_307_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_307_dilations_0 = const()[name = string("x_307_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239819776))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239829056))))[name = string("encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_47_cast_fp16)[name = string("x_307_cast_fp16")];
+            tensor<int32, [3]> input_627_perm_0 = const()[name = string("input_627_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_309_axes_0 = const()[name = string("x_309_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239831168)))];
+            tensor<fp16, [1024]> encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239833280)))];
+            tensor<fp16, [1, 28, 1024]> input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = string("transpose_257")];
+            tensor<fp16, [1, 28, 1024]> x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = string("x_309_cast_fp16")];
+            tensor<int32, [3]> input_629_perm_0 = const()[name = string("input_629_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = string("transpose_256")];
+            tensor<fp16, [1, 1024, 28]> input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = string("input_631_cast_fp16")];
+            string x_311_pad_type_0 = const()[name = string("x_311_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_311_strides_0 = const()[name = string("x_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_311_pad_0 = const()[name = string("x_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_311_dilations_0 = const()[name = string("x_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_311_groups_0 = const()[name = string("x_311_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239835392))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240884032))))[name = string("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_311_cast_fp16 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_631_cast_fp16)[name = string("x_311_cast_fp16")];
+            tensor<int32, [3]> input_633_perm_0 = const()[name = string("input_633_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = string("transpose_255")];
+            tensor<fp16, [1, 28, 1024]> input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = string("input_635_cast_fp16")];
+            tensor<int32, [1]> input_637_axes_0 = const()[name = string("input_637_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240886144)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240888256)))];
+            tensor<fp16, [1, 28, 1024]> input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = string("input_637_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240890368))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244036160))))[name = string("encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_11_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244036352)))];
+            tensor<fp16, [1, 28, 4096]> linear_107_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized, x = input_637_cast_fp16)[name = string("linear_107_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = string("input_641_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244044608))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247190400))))[name = string("encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247190592)))];
+            tensor<fp16, [1, 28, 1024]> linear_108_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized, x = input_641_cast_fp16)[name = string("linear_108_cast_fp16")];
+            fp16 var_2974_to_fp16 = const()[name = string("op_2974_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_2975_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2974_to_fp16)[name = string("op_2975_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2975_cast_fp16)[name = string("input_647_cast_fp16")];
+            tensor<int32, [1]> input_649_axes_0 = const()[name = string("input_649_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247192704)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247194816)))];
+            tensor<fp16, [1, 28, 1024]> input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = string("input_649_cast_fp16")];
+            tensor<int32, [4]> cache_49_begin_0 = const()[name = string("cache_49_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_49_end_0 = const()[name = string("cache_49_end_0"), val = tensor<int32, [4]>([13, 1, 42, 1024])];
+            tensor<bool, [4]> cache_49_end_mask_0 = const()[name = string("cache_49_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_49_squeeze_mask_0 = const()[name = string("cache_49_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_49_cast_fp16 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_49_cast_fp16")];
+            tensor<int32, [4]> cache_51_begin_0 = const()[name = string("cache_51_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_51_end_0 = const()[name = string("cache_51_end_0"), val = tensor<int32, [4]>([13, 1, 1024, 8])];
+            tensor<bool, [4]> cache_51_end_mask_0 = const()[name = string("cache_51_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_51_squeeze_mask_0 = const()[name = string("cache_51_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_51_cast_fp16 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_51_cast_fp16")];
+            tensor<int32, [1]> input_651_axes_0 = const()[name = string("input_651_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247196928)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247199040)))];
+            tensor<fp16, [1, 28, 1024]> input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = string("input_651_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247201152))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250346944))))[name = string("encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_12_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250347136)))];
+            tensor<fp16, [1, 28, 4096]> linear_109_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized, x = input_651_cast_fp16)[name = string("linear_109_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("input_655_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(250355392))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253501184))))[name = string("encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253501376)))];
+            tensor<fp16, [1, 28, 1024]> linear_110_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized, x = input_655_cast_fp16)[name = string("linear_110_cast_fp16")];
+            fp16 var_3011_to_fp16 = const()[name = string("op_3011_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_3012_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_3011_to_fp16)[name = string("op_3012_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_3012_cast_fp16)[name = string("input_661_cast_fp16")];
+            tensor<int32, [1]> key_25_axes_0 = const()[name = string("key_25_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253503488)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253505600)))];
+            tensor<fp16, [1, 28, 1024]> key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = string("key_25_cast_fp16")];
+            bool input_663_interleave_0 = const()[name = string("input_663_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_663_cast_fp16 = concat(axis = var_68, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = string("input_663_cast_fp16")];
+            tensor<int32, [3]> var_3034_begin_0 = const()[name = string("op_3034_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_3034_end_0 = const()[name = string("op_3034_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3034_end_mask_0 = const()[name = string("op_3034_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_3034_cast_fp16 = slice_by_index(begin = var_3034_begin_0, end = var_3034_end_0, end_mask = var_3034_end_mask_0, x = cache_49_cast_fp16)[name = string("op_3034_cast_fp16")];
+            bool var_3040_interleave_0 = const()[name = string("op_3040_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3040_cast_fp16 = concat(axis = var_68, interleave = var_3040_interleave_0, values = (var_3034_cast_fp16, key_25_cast_fp16))[name = string("op_3040_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253507712))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254294208))))[name = string("encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254294400)))];
+            tensor<fp16, [1, 28, 1024]> linear_111_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized, x = key_25_cast_fp16)[name = string("linear_111_cast_fp16")];
+            tensor<int32, [4]> var_3045 = const()[name = string("op_3045"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_73_cast_fp16 = reshape(shape = var_3045, x = linear_111_cast_fp16)[name = string("q_73_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254296512))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255083008))))[name = string("encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255083200)))];
+            tensor<fp16, [1, 70, 1024]> linear_112_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_112_cast_fp16")];
+            tensor<int32, [4]> var_3050 = const()[name = string("op_3050"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_49_cast_fp16 = reshape(shape = var_3050, x = linear_112_cast_fp16)[name = string("k_49_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255085312))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255871808))))[name = string("encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255872000)))];
+            tensor<fp16, [1, 70, 1024]> linear_113_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_113_cast_fp16")];
+            tensor<int32, [4]> var_3055 = const()[name = string("op_3055"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_25_cast_fp16 = reshape(shape = var_3055, x = linear_113_cast_fp16)[name = string("v_25_cast_fp16")];
+            tensor<int32, [4]> value_33_perm_0 = const()[name = string("value_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255874112)))];
+            tensor<fp16, [1, 28, 8, 128]> var_3068_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = string("op_3068_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255876224)))];
+            tensor<fp16, [1, 28, 8, 128]> var_3070_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = string("op_3070_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_25_perm_0 = const()[name = string("q_with_bias_v_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_319_transpose_x_0 = const()[name = string("x_319_transpose_x_0"), val = bool(false)];
+            bool x_319_transpose_y_0 = const()[name = string("x_319_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_3072_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255878336))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256020736))))[name = string("op_3072_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_3070_cast_fp16)[name = string("transpose_254")];
+            tensor<fp16, [1, 8, 28, 139]> x_319_cast_fp16 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = op_3072_to_fp16_quantized)[name = string("x_319_cast_fp16")];
+            tensor<int32, [8]> x_321_pad_0 = const()[name = string("x_321_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("constant")];
+            fp16 const_235_to_fp16 = const()[name = string("const_235_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_321_cast_fp16 = pad(constant_val = const_235_to_fp16, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319_cast_fp16)[name = string("x_321_cast_fp16")];
+            tensor<int32, [4]> var_3080 = const()[name = string("op_3080"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_323_cast_fp16 = reshape(shape = var_3080, x = x_321_cast_fp16)[name = string("x_323_cast_fp16")];
+            tensor<int32, [4]> var_3084_begin_0 = const()[name = string("op_3084_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3084_end_0 = const()[name = string("op_3084_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_3084_end_mask_0 = const()[name = string("op_3084_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_3084_cast_fp16 = slice_by_index(begin = var_3084_begin_0, end = var_3084_end_0, end_mask = var_3084_end_mask_0, x = x_323_cast_fp16)[name = string("op_3084_cast_fp16")];
+            tensor<int32, [4]> var_3085 = const()[name = string("op_3085"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_49_cast_fp16 = reshape(shape = var_3085, x = var_3084_cast_fp16)[name = string("matrix_bd_49_cast_fp16")];
+            bool matrix_ac_25_transpose_x_0 = const()[name = string("matrix_ac_25_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_25_transpose_y_0 = const()[name = string("matrix_ac_25_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_120_perm_0 = const()[name = string("transpose_120_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_121_perm_0 = const()[name = string("transpose_121_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_121 = transpose(perm = transpose_121_perm_0, x = k_49_cast_fp16)[name = string("transpose_252")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_120 = transpose(perm = transpose_120_perm_0, x = var_3068_cast_fp16)[name = string("transpose_253")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_25_cast_fp16 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_120, y = transpose_121)[name = string("matrix_ac_25_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_51_begin_0 = const()[name = string("matrix_bd_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_51_end_0 = const()[name = string("matrix_bd_51_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_51_end_mask_0 = const()[name = string("matrix_bd_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_51_cast_fp16 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49_cast_fp16)[name = string("matrix_bd_51_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_3094_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = string("op_3094_cast_fp16")];
+            fp16 _inversed_scores_49_y_0_to_fp16 = const()[name = string("_inversed_scores_49_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_49_cast_fp16 = mul(x = var_3094_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = string("_inversed_scores_49_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_51_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_11)[name = string("scores_51_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_3100_cast_fp16 = softmax(axis = var_59, x = scores_51_cast_fp16)[name = string("op_3100_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_665_cast_fp16 = select(a = var_44_to_fp16, b = var_3100_cast_fp16, cond = mask_11)[name = string("input_665_cast_fp16")];
+            bool x_325_transpose_x_0 = const()[name = string("x_325_transpose_x_0"), val = bool(false)];
+            bool x_325_transpose_y_0 = const()[name = string("x_325_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_25_cast_fp16)[name = string("transpose_251")];
+            tensor<fp16, [1, 8, 28, 128]> x_325_cast_fp16 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665_cast_fp16, y = value_33_cast_fp16)[name = string("x_325_cast_fp16")];
+            tensor<int32, [4]> var_3104_perm_0 = const()[name = string("op_3104_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3105 = const()[name = string("op_3105"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_3104_cast_fp16 = transpose(perm = var_3104_perm_0, x = x_325_cast_fp16)[name = string("transpose_250")];
+            tensor<fp16, [1, 28, 1024]> input_667_cast_fp16 = reshape(shape = var_3105, x = var_3104_cast_fp16)[name = string("input_667_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256021120))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256807616))))[name = string("encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256807808)))];
+            tensor<fp16, [1, 28, 1024]> linear_115_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized, x = input_667_cast_fp16)[name = string("linear_115_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = string("input_671_cast_fp16")];
+            tensor<int32, [1]> x_329_axes_0 = const()[name = string("x_329_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256809920)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256812032)))];
+            tensor<fp16, [1, 28, 1024]> x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = string("x_329_cast_fp16")];
+            tensor<int32, [3]> input_673_perm_0 = const()[name = string("input_673_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_675_pad_type_0 = const()[name = string("input_675_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_675_strides_0 = const()[name = string("input_675_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_675_pad_0 = const()[name = string("input_675_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_675_dilations_0 = const()[name = string("input_675_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_675_groups_0 = const()[name = string("input_675_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256814144))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258911360))))[name = string("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = string("transpose_249")];
+            tensor<fp16, [1, 2048, 28]> input_675_cast_fp16 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_673_cast_fp16)[name = string("input_675_cast_fp16")];
+            int32 x_331_split_num_splits_0 = const()[name = string("x_331_split_num_splits_0"), val = int32(2)];
+            int32 x_331_split_axis_0 = const()[name = string("x_331_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_331_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_331_split_cast_fp16_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675_cast_fp16)[name = string("x_331_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = string("x_331_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = string("x_331_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_677_cast_fp16 = select(a = var_44_to_fp16, b = x_331_cast_fp16, cond = var_575)[name = string("input_677_cast_fp16")];
+            bool new_x_51_interleave_0 = const()[name = string("new_x_51_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_51_cast_fp16 = concat(axis = var_59, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = string("new_x_51_cast_fp16")];
+            tensor<int32, [3]> var_3144_begin_0 = const()[name = string("op_3144_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_3144_end_0 = const()[name = string("op_3144_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_3144_end_mask_0 = const()[name = string("op_3144_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3144_cast_fp16 = slice_by_index(begin = var_3144_begin_0, end = var_3144_end_0, end_mask = var_3144_end_mask_0, x = new_x_51_cast_fp16)[name = string("op_3144_cast_fp16")];
+            string x_333_pad_type_0 = const()[name = string("x_333_pad_type_0"), val = string("valid")];
+            int32 x_333_groups_0 = const()[name = string("x_333_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_333_strides_0 = const()[name = string("x_333_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_333_pad_0 = const()[name = string("x_333_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_333_dilations_0 = const()[name = string("x_333_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258915520))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258924800))))[name = string("encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_333_cast_fp16 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_51_cast_fp16)[name = string("x_333_cast_fp16")];
+            tensor<int32, [3]> input_679_perm_0 = const()[name = string("input_679_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_335_axes_0 = const()[name = string("x_335_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258926912)))];
+            tensor<fp16, [1024]> encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258929024)))];
+            tensor<fp16, [1, 28, 1024]> input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = string("transpose_248")];
+            tensor<fp16, [1, 28, 1024]> x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = string("x_335_cast_fp16")];
+            tensor<int32, [3]> input_681_perm_0 = const()[name = string("input_681_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = string("transpose_247")];
+            tensor<fp16, [1, 1024, 28]> input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = string("input_683_cast_fp16")];
+            string x_337_pad_type_0 = const()[name = string("x_337_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_337_strides_0 = const()[name = string("x_337_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_337_pad_0 = const()[name = string("x_337_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_337_dilations_0 = const()[name = string("x_337_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_337_groups_0 = const()[name = string("x_337_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258931136))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259979776))))[name = string("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_337_cast_fp16 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_683_cast_fp16)[name = string("x_337_cast_fp16")];
+            tensor<int32, [3]> input_685_perm_0 = const()[name = string("input_685_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = string("transpose_246")];
+            tensor<fp16, [1, 28, 1024]> input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = string("input_687_cast_fp16")];
+            tensor<int32, [1]> input_689_axes_0 = const()[name = string("input_689_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259981888)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259984000)))];
+            tensor<fp16, [1, 28, 1024]> input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = string("input_689_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259986112))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263131904))))[name = string("encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_12_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263132096)))];
+            tensor<fp16, [1, 28, 4096]> linear_116_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized, x = input_689_cast_fp16)[name = string("linear_116_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("input_693_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263140352))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266286144))))[name = string("encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266286336)))];
+            tensor<fp16, [1, 28, 1024]> linear_117_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized, x = input_693_cast_fp16)[name = string("linear_117_cast_fp16")];
+            fp16 var_3187_to_fp16 = const()[name = string("op_3187_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_3188_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_3187_to_fp16)[name = string("op_3188_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_3188_cast_fp16)[name = string("input_699_cast_fp16")];
+            tensor<int32, [1]> input_701_axes_0 = const()[name = string("input_701_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266288448)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266290560)))];
+            tensor<fp16, [1, 28, 1024]> input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = string("input_701_cast_fp16")];
+            tensor<int32, [4]> cache_53_begin_0 = const()[name = string("cache_53_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_53_end_0 = const()[name = string("cache_53_end_0"), val = tensor<int32, [4]>([14, 1, 42, 1024])];
+            tensor<bool, [4]> cache_53_end_mask_0 = const()[name = string("cache_53_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_53_squeeze_mask_0 = const()[name = string("cache_53_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_53_cast_fp16 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_53_cast_fp16")];
+            tensor<int32, [4]> cache_55_begin_0 = const()[name = string("cache_55_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_55_end_0 = const()[name = string("cache_55_end_0"), val = tensor<int32, [4]>([14, 1, 1024, 8])];
+            tensor<bool, [4]> cache_55_end_mask_0 = const()[name = string("cache_55_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_55_squeeze_mask_0 = const()[name = string("cache_55_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_55_cast_fp16 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_55_cast_fp16")];
+            tensor<int32, [1]> input_703_axes_0 = const()[name = string("input_703_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266292672)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266294784)))];
+            tensor<fp16, [1, 28, 1024]> input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = string("input_703_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266296896))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269442688))))[name = string("encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_13_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269442880)))];
+            tensor<fp16, [1, 28, 4096]> linear_118_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized, x = input_703_cast_fp16)[name = string("linear_118_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = string("input_707_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(269451136))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272596928))))[name = string("encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272597120)))];
+            tensor<fp16, [1, 28, 1024]> linear_119_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized, x = input_707_cast_fp16)[name = string("linear_119_cast_fp16")];
+            fp16 var_3224_to_fp16 = const()[name = string("op_3224_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_3225_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_3224_to_fp16)[name = string("op_3225_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_3225_cast_fp16)[name = string("input_713_cast_fp16")];
+            tensor<int32, [1]> key_27_axes_0 = const()[name = string("key_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272599232)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272601344)))];
+            tensor<fp16, [1, 28, 1024]> key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = string("key_27_cast_fp16")];
+            bool input_715_interleave_0 = const()[name = string("input_715_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_715_cast_fp16 = concat(axis = var_68, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = string("input_715_cast_fp16")];
+            tensor<int32, [3]> var_3247_begin_0 = const()[name = string("op_3247_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_3247_end_0 = const()[name = string("op_3247_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3247_end_mask_0 = const()[name = string("op_3247_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_3247_cast_fp16 = slice_by_index(begin = var_3247_begin_0, end = var_3247_end_0, end_mask = var_3247_end_mask_0, x = cache_53_cast_fp16)[name = string("op_3247_cast_fp16")];
+            bool var_3253_interleave_0 = const()[name = string("op_3253_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3253_cast_fp16 = concat(axis = var_68, interleave = var_3253_interleave_0, values = (var_3247_cast_fp16, key_27_cast_fp16))[name = string("op_3253_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272603456))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273389952))))[name = string("encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273390144)))];
+            tensor<fp16, [1, 28, 1024]> linear_120_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized, x = key_27_cast_fp16)[name = string("linear_120_cast_fp16")];
+            tensor<int32, [4]> var_3258 = const()[name = string("op_3258"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_79_cast_fp16 = reshape(shape = var_3258, x = linear_120_cast_fp16)[name = string("q_79_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273392256))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274178752))))[name = string("encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274178944)))];
+            tensor<fp16, [1, 70, 1024]> linear_121_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_121_cast_fp16")];
+            tensor<int32, [4]> var_3263 = const()[name = string("op_3263"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_53_cast_fp16 = reshape(shape = var_3263, x = linear_121_cast_fp16)[name = string("k_53_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274181056))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274967552))))[name = string("encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274967744)))];
+            tensor<fp16, [1, 70, 1024]> linear_122_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_122_cast_fp16")];
+            tensor<int32, [4]> var_3268 = const()[name = string("op_3268"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_27_cast_fp16 = reshape(shape = var_3268, x = linear_122_cast_fp16)[name = string("v_27_cast_fp16")];
+            tensor<int32, [4]> value_35_perm_0 = const()[name = string("value_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274969856)))];
+            tensor<fp16, [1, 28, 8, 128]> var_3281_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = string("op_3281_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274971968)))];
+            tensor<fp16, [1, 28, 8, 128]> var_3283_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = string("op_3283_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_27_perm_0 = const()[name = string("q_with_bias_v_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_345_transpose_x_0 = const()[name = string("x_345_transpose_x_0"), val = bool(false)];
+            bool x_345_transpose_y_0 = const()[name = string("x_345_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_3285_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274974080))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275116480))))[name = string("op_3285_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3283_cast_fp16)[name = string("transpose_245")];
+            tensor<fp16, [1, 8, 28, 139]> x_345_cast_fp16 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = op_3285_to_fp16_quantized)[name = string("x_345_cast_fp16")];
+            tensor<int32, [8]> x_347_pad_0 = const()[name = string("x_347_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_347_mode_0 = const()[name = string("x_347_mode_0"), val = string("constant")];
+            fp16 const_248_to_fp16 = const()[name = string("const_248_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_347_cast_fp16 = pad(constant_val = const_248_to_fp16, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345_cast_fp16)[name = string("x_347_cast_fp16")];
+            tensor<int32, [4]> var_3293 = const()[name = string("op_3293"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_349_cast_fp16 = reshape(shape = var_3293, x = x_347_cast_fp16)[name = string("x_349_cast_fp16")];
+            tensor<int32, [4]> var_3297_begin_0 = const()[name = string("op_3297_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3297_end_0 = const()[name = string("op_3297_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_3297_end_mask_0 = const()[name = string("op_3297_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_3297_cast_fp16 = slice_by_index(begin = var_3297_begin_0, end = var_3297_end_0, end_mask = var_3297_end_mask_0, x = x_349_cast_fp16)[name = string("op_3297_cast_fp16")];
+            tensor<int32, [4]> var_3298 = const()[name = string("op_3298"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_53_cast_fp16 = reshape(shape = var_3298, x = var_3297_cast_fp16)[name = string("matrix_bd_53_cast_fp16")];
+            bool matrix_ac_27_transpose_x_0 = const()[name = string("matrix_ac_27_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_27_transpose_y_0 = const()[name = string("matrix_ac_27_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_122_perm_0 = const()[name = string("transpose_122_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_123_perm_0 = const()[name = string("transpose_123_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_123 = transpose(perm = transpose_123_perm_0, x = k_53_cast_fp16)[name = string("transpose_243")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_122 = transpose(perm = transpose_122_perm_0, x = var_3281_cast_fp16)[name = string("transpose_244")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_27_cast_fp16 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_122, y = transpose_123)[name = string("matrix_ac_27_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_55_begin_0 = const()[name = string("matrix_bd_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_55_end_0 = const()[name = string("matrix_bd_55_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_55_end_mask_0 = const()[name = string("matrix_bd_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_55_cast_fp16 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53_cast_fp16)[name = string("matrix_bd_55_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_3307_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = string("op_3307_cast_fp16")];
+            fp16 _inversed_scores_53_y_0_to_fp16 = const()[name = string("_inversed_scores_53_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_53_cast_fp16 = mul(x = var_3307_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = string("_inversed_scores_53_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_55_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_11)[name = string("scores_55_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_3313_cast_fp16 = softmax(axis = var_59, x = scores_55_cast_fp16)[name = string("op_3313_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_717_cast_fp16 = select(a = var_44_to_fp16, b = var_3313_cast_fp16, cond = mask_11)[name = string("input_717_cast_fp16")];
+            bool x_351_transpose_x_0 = const()[name = string("x_351_transpose_x_0"), val = bool(false)];
+            bool x_351_transpose_y_0 = const()[name = string("x_351_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = v_27_cast_fp16)[name = string("transpose_242")];
+            tensor<fp16, [1, 8, 28, 128]> x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717_cast_fp16, y = value_35_cast_fp16)[name = string("x_351_cast_fp16")];
+            tensor<int32, [4]> var_3317_perm_0 = const()[name = string("op_3317_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3318 = const()[name = string("op_3318"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_3317_cast_fp16 = transpose(perm = var_3317_perm_0, x = x_351_cast_fp16)[name = string("transpose_241")];
+            tensor<fp16, [1, 28, 1024]> input_719_cast_fp16 = reshape(shape = var_3318, x = var_3317_cast_fp16)[name = string("input_719_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275116864))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275903360))))[name = string("encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275903552)))];
+            tensor<fp16, [1, 28, 1024]> linear_124_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("linear_124_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = string("input_723_cast_fp16")];
+            tensor<int32, [1]> x_355_axes_0 = const()[name = string("x_355_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275905664)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275907776)))];
+            tensor<fp16, [1, 28, 1024]> x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = string("x_355_cast_fp16")];
+            tensor<int32, [3]> input_725_perm_0 = const()[name = string("input_725_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_727_pad_type_0 = const()[name = string("input_727_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_727_strides_0 = const()[name = string("input_727_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_727_pad_0 = const()[name = string("input_727_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_727_dilations_0 = const()[name = string("input_727_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_727_groups_0 = const()[name = string("input_727_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275909888))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278007104))))[name = string("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = string("transpose_240")];
+            tensor<fp16, [1, 2048, 28]> input_727_cast_fp16 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_725_cast_fp16)[name = string("input_727_cast_fp16")];
+            int32 x_357_split_num_splits_0 = const()[name = string("x_357_split_num_splits_0"), val = int32(2)];
+            int32 x_357_split_axis_0 = const()[name = string("x_357_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_357_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727_cast_fp16)[name = string("x_357_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = string("x_357_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = string("x_357_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_729_cast_fp16 = select(a = var_44_to_fp16, b = x_357_cast_fp16, cond = var_575)[name = string("input_729_cast_fp16")];
+            bool new_x_55_interleave_0 = const()[name = string("new_x_55_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_55_cast_fp16 = concat(axis = var_59, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = string("new_x_55_cast_fp16")];
+            tensor<int32, [3]> var_3357_begin_0 = const()[name = string("op_3357_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_3357_end_0 = const()[name = string("op_3357_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_3357_end_mask_0 = const()[name = string("op_3357_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3357_cast_fp16 = slice_by_index(begin = var_3357_begin_0, end = var_3357_end_0, end_mask = var_3357_end_mask_0, x = new_x_55_cast_fp16)[name = string("op_3357_cast_fp16")];
+            string x_359_pad_type_0 = const()[name = string("x_359_pad_type_0"), val = string("valid")];
+            int32 x_359_groups_0 = const()[name = string("x_359_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_359_strides_0 = const()[name = string("x_359_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_359_pad_0 = const()[name = string("x_359_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_359_dilations_0 = const()[name = string("x_359_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278011264))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278020544))))[name = string("encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_55_cast_fp16)[name = string("x_359_cast_fp16")];
+            tensor<int32, [3]> input_731_perm_0 = const()[name = string("input_731_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_361_axes_0 = const()[name = string("x_361_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278022656)))];
+            tensor<fp16, [1024]> encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278024768)))];
+            tensor<fp16, [1, 28, 1024]> input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = string("transpose_239")];
+            tensor<fp16, [1, 28, 1024]> x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = string("x_361_cast_fp16")];
+            tensor<int32, [3]> input_733_perm_0 = const()[name = string("input_733_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = string("transpose_238")];
+            tensor<fp16, [1, 1024, 28]> input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = string("input_735_cast_fp16")];
+            string x_363_pad_type_0 = const()[name = string("x_363_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_363_strides_0 = const()[name = string("x_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_363_pad_0 = const()[name = string("x_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_363_dilations_0 = const()[name = string("x_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_363_groups_0 = const()[name = string("x_363_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278026880))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279075520))))[name = string("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_735_cast_fp16)[name = string("x_363_cast_fp16")];
+            tensor<int32, [3]> input_737_perm_0 = const()[name = string("input_737_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = string("transpose_237")];
+            tensor<fp16, [1, 28, 1024]> input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = string("input_739_cast_fp16")];
+            tensor<int32, [1]> input_741_axes_0 = const()[name = string("input_741_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279077632)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279079744)))];
+            tensor<fp16, [1, 28, 1024]> input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = string("input_741_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279081856))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282227648))))[name = string("encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_13_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282227840)))];
+            tensor<fp16, [1, 28, 4096]> linear_125_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized, x = input_741_cast_fp16)[name = string("linear_125_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = string("input_745_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282236096))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285381888))))[name = string("encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285382080)))];
+            tensor<fp16, [1, 28, 1024]> linear_126_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized, x = input_745_cast_fp16)[name = string("linear_126_cast_fp16")];
+            fp16 var_3400_to_fp16 = const()[name = string("op_3400_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_3401_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3400_to_fp16)[name = string("op_3401_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3401_cast_fp16)[name = string("input_751_cast_fp16")];
+            tensor<int32, [1]> input_753_axes_0 = const()[name = string("input_753_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285384192)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285386304)))];
+            tensor<fp16, [1, 28, 1024]> input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = string("input_753_cast_fp16")];
+            tensor<int32, [4]> cache_57_begin_0 = const()[name = string("cache_57_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_57_end_0 = const()[name = string("cache_57_end_0"), val = tensor<int32, [4]>([15, 1, 42, 1024])];
+            tensor<bool, [4]> cache_57_end_mask_0 = const()[name = string("cache_57_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_57_squeeze_mask_0 = const()[name = string("cache_57_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_57_cast_fp16 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_57_cast_fp16")];
+            tensor<int32, [4]> cache_59_begin_0 = const()[name = string("cache_59_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_59_end_0 = const()[name = string("cache_59_end_0"), val = tensor<int32, [4]>([15, 1, 1024, 8])];
+            tensor<bool, [4]> cache_59_end_mask_0 = const()[name = string("cache_59_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_59_squeeze_mask_0 = const()[name = string("cache_59_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_59_cast_fp16 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_59_cast_fp16")];
+            tensor<int32, [1]> input_755_axes_0 = const()[name = string("input_755_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285388416)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285390528)))];
+            tensor<fp16, [1, 28, 1024]> input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = string("input_755_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285392640))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288538432))))[name = string("encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_14_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288538624)))];
+            tensor<fp16, [1, 28, 4096]> linear_127_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized, x = input_755_cast_fp16)[name = string("linear_127_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = string("input_759_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(288546880))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291692672))))[name = string("encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291692864)))];
+            tensor<fp16, [1, 28, 1024]> linear_128_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized, x = input_759_cast_fp16)[name = string("linear_128_cast_fp16")];
+            fp16 var_3437_to_fp16 = const()[name = string("op_3437_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_3438_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3437_to_fp16)[name = string("op_3438_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3438_cast_fp16)[name = string("input_765_cast_fp16")];
+            tensor<int32, [1]> key_29_axes_0 = const()[name = string("key_29_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291694976)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291697088)))];
+            tensor<fp16, [1, 28, 1024]> key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = string("key_29_cast_fp16")];
+            bool input_767_interleave_0 = const()[name = string("input_767_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_767_cast_fp16 = concat(axis = var_68, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = string("input_767_cast_fp16")];
+            tensor<int32, [3]> var_3460_begin_0 = const()[name = string("op_3460_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_3460_end_0 = const()[name = string("op_3460_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3460_end_mask_0 = const()[name = string("op_3460_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_3460_cast_fp16 = slice_by_index(begin = var_3460_begin_0, end = var_3460_end_0, end_mask = var_3460_end_mask_0, x = cache_57_cast_fp16)[name = string("op_3460_cast_fp16")];
+            bool var_3466_interleave_0 = const()[name = string("op_3466_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3466_cast_fp16 = concat(axis = var_68, interleave = var_3466_interleave_0, values = (var_3460_cast_fp16, key_29_cast_fp16))[name = string("op_3466_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291699200))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292485696))))[name = string("encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292485888)))];
+            tensor<fp16, [1, 28, 1024]> linear_129_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized, x = key_29_cast_fp16)[name = string("linear_129_cast_fp16")];
+            tensor<int32, [4]> var_3471 = const()[name = string("op_3471"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_85_cast_fp16 = reshape(shape = var_3471, x = linear_129_cast_fp16)[name = string("q_85_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292488000))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293274496))))[name = string("encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293274688)))];
+            tensor<fp16, [1, 70, 1024]> linear_130_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_130_cast_fp16")];
+            tensor<int32, [4]> var_3476 = const()[name = string("op_3476"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_57_cast_fp16 = reshape(shape = var_3476, x = linear_130_cast_fp16)[name = string("k_57_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293276800))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294063296))))[name = string("encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294063488)))];
+            tensor<fp16, [1, 70, 1024]> linear_131_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_131_cast_fp16")];
+            tensor<int32, [4]> var_3481 = const()[name = string("op_3481"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_29_cast_fp16 = reshape(shape = var_3481, x = linear_131_cast_fp16)[name = string("v_29_cast_fp16")];
+            tensor<int32, [4]> value_37_perm_0 = const()[name = string("value_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294065600)))];
+            tensor<fp16, [1, 28, 8, 128]> var_3494_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = string("op_3494_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294067712)))];
+            tensor<fp16, [1, 28, 8, 128]> var_3496_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = string("op_3496_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_29_perm_0 = const()[name = string("q_with_bias_v_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_371_transpose_x_0 = const()[name = string("x_371_transpose_x_0"), val = bool(false)];
+            bool x_371_transpose_y_0 = const()[name = string("x_371_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_3498_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294069824))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294212224))))[name = string("op_3498_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3496_cast_fp16)[name = string("transpose_236")];
+            tensor<fp16, [1, 8, 28, 139]> x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = op_3498_to_fp16_quantized)[name = string("x_371_cast_fp16")];
+            tensor<int32, [8]> x_373_pad_0 = const()[name = string("x_373_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_373_mode_0 = const()[name = string("x_373_mode_0"), val = string("constant")];
+            fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_373_cast_fp16 = pad(constant_val = const_261_to_fp16, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371_cast_fp16)[name = string("x_373_cast_fp16")];
+            tensor<int32, [4]> var_3506 = const()[name = string("op_3506"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_375_cast_fp16 = reshape(shape = var_3506, x = x_373_cast_fp16)[name = string("x_375_cast_fp16")];
+            tensor<int32, [4]> var_3510_begin_0 = const()[name = string("op_3510_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3510_end_0 = const()[name = string("op_3510_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_3510_end_mask_0 = const()[name = string("op_3510_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_3510_cast_fp16 = slice_by_index(begin = var_3510_begin_0, end = var_3510_end_0, end_mask = var_3510_end_mask_0, x = x_375_cast_fp16)[name = string("op_3510_cast_fp16")];
+            tensor<int32, [4]> var_3511 = const()[name = string("op_3511"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_57_cast_fp16 = reshape(shape = var_3511, x = var_3510_cast_fp16)[name = string("matrix_bd_57_cast_fp16")];
+            bool matrix_ac_29_transpose_x_0 = const()[name = string("matrix_ac_29_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_29_transpose_y_0 = const()[name = string("matrix_ac_29_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_124_perm_0 = const()[name = string("transpose_124_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_125_perm_0 = const()[name = string("transpose_125_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_125 = transpose(perm = transpose_125_perm_0, x = k_57_cast_fp16)[name = string("transpose_234")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_124 = transpose(perm = transpose_124_perm_0, x = var_3494_cast_fp16)[name = string("transpose_235")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_29_cast_fp16 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_124, y = transpose_125)[name = string("matrix_ac_29_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_59_begin_0 = const()[name = string("matrix_bd_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_59_end_0 = const()[name = string("matrix_bd_59_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_59_end_mask_0 = const()[name = string("matrix_bd_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_59_cast_fp16 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57_cast_fp16)[name = string("matrix_bd_59_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_3520_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = string("op_3520_cast_fp16")];
+            fp16 _inversed_scores_57_y_0_to_fp16 = const()[name = string("_inversed_scores_57_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_57_cast_fp16 = mul(x = var_3520_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = string("_inversed_scores_57_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_59_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_11)[name = string("scores_59_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_3526_cast_fp16 = softmax(axis = var_59, x = scores_59_cast_fp16)[name = string("op_3526_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_769_cast_fp16 = select(a = var_44_to_fp16, b = var_3526_cast_fp16, cond = mask_11)[name = string("input_769_cast_fp16")];
+            bool x_377_transpose_x_0 = const()[name = string("x_377_transpose_x_0"), val = bool(false)];
+            bool x_377_transpose_y_0 = const()[name = string("x_377_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = v_29_cast_fp16)[name = string("transpose_233")];
+            tensor<fp16, [1, 8, 28, 128]> x_377_cast_fp16 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769_cast_fp16, y = value_37_cast_fp16)[name = string("x_377_cast_fp16")];
+            tensor<int32, [4]> var_3530_perm_0 = const()[name = string("op_3530_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3531 = const()[name = string("op_3531"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_3530_cast_fp16 = transpose(perm = var_3530_perm_0, x = x_377_cast_fp16)[name = string("transpose_232")];
+            tensor<fp16, [1, 28, 1024]> input_771_cast_fp16 = reshape(shape = var_3531, x = var_3530_cast_fp16)[name = string("input_771_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294212608))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294999104))))[name = string("encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294999296)))];
+            tensor<fp16, [1, 28, 1024]> linear_133_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized, x = input_771_cast_fp16)[name = string("linear_133_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = string("input_775_cast_fp16")];
+            tensor<int32, [1]> x_381_axes_0 = const()[name = string("x_381_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295001408)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295003520)))];
+            tensor<fp16, [1, 28, 1024]> x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = string("x_381_cast_fp16")];
+            tensor<int32, [3]> input_777_perm_0 = const()[name = string("input_777_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_779_pad_type_0 = const()[name = string("input_779_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_779_strides_0 = const()[name = string("input_779_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_779_pad_0 = const()[name = string("input_779_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_779_dilations_0 = const()[name = string("input_779_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_779_groups_0 = const()[name = string("input_779_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295005632))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297102848))))[name = string("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = string("transpose_231")];
+            tensor<fp16, [1, 2048, 28]> input_779_cast_fp16 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_777_cast_fp16)[name = string("input_779_cast_fp16")];
+            int32 x_383_split_num_splits_0 = const()[name = string("x_383_split_num_splits_0"), val = int32(2)];
+            int32 x_383_split_axis_0 = const()[name = string("x_383_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_383_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_383_split_cast_fp16_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779_cast_fp16)[name = string("x_383_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = string("x_383_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = string("x_383_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_781_cast_fp16 = select(a = var_44_to_fp16, b = x_383_cast_fp16, cond = var_575)[name = string("input_781_cast_fp16")];
+            bool new_x_59_interleave_0 = const()[name = string("new_x_59_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_59_cast_fp16 = concat(axis = var_59, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = string("new_x_59_cast_fp16")];
+            tensor<int32, [3]> var_3570_begin_0 = const()[name = string("op_3570_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_3570_end_0 = const()[name = string("op_3570_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_3570_end_mask_0 = const()[name = string("op_3570_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3570_cast_fp16 = slice_by_index(begin = var_3570_begin_0, end = var_3570_end_0, end_mask = var_3570_end_mask_0, x = new_x_59_cast_fp16)[name = string("op_3570_cast_fp16")];
+            string x_385_pad_type_0 = const()[name = string("x_385_pad_type_0"), val = string("valid")];
+            int32 x_385_groups_0 = const()[name = string("x_385_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_385_strides_0 = const()[name = string("x_385_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_385_pad_0 = const()[name = string("x_385_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_385_dilations_0 = const()[name = string("x_385_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297107008))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297116288))))[name = string("encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_59_cast_fp16)[name = string("x_385_cast_fp16")];
+            tensor<int32, [3]> input_783_perm_0 = const()[name = string("input_783_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_387_axes_0 = const()[name = string("x_387_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297118400)))];
+            tensor<fp16, [1024]> encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297120512)))];
+            tensor<fp16, [1, 28, 1024]> input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = string("transpose_230")];
+            tensor<fp16, [1, 28, 1024]> x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = string("x_387_cast_fp16")];
+            tensor<int32, [3]> input_785_perm_0 = const()[name = string("input_785_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = string("transpose_229")];
+            tensor<fp16, [1, 1024, 28]> input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = string("input_787_cast_fp16")];
+            string x_389_pad_type_0 = const()[name = string("x_389_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_389_strides_0 = const()[name = string("x_389_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_389_pad_0 = const()[name = string("x_389_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_389_dilations_0 = const()[name = string("x_389_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_389_groups_0 = const()[name = string("x_389_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297122624))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298171264))))[name = string("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_389_cast_fp16 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_787_cast_fp16)[name = string("x_389_cast_fp16")];
+            tensor<int32, [3]> input_789_perm_0 = const()[name = string("input_789_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = string("transpose_228")];
+            tensor<fp16, [1, 28, 1024]> input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = string("input_791_cast_fp16")];
+            tensor<int32, [1]> input_793_axes_0 = const()[name = string("input_793_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298173376)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298175488)))];
+            tensor<fp16, [1, 28, 1024]> input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = string("input_793_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298177600))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301323392))))[name = string("encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_14_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301323584)))];
+            tensor<fp16, [1, 28, 4096]> linear_134_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized, x = input_793_cast_fp16)[name = string("linear_134_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = string("input_797_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301331840))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304477632))))[name = string("encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304477824)))];
+            tensor<fp16, [1, 28, 1024]> linear_135_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized, x = input_797_cast_fp16)[name = string("linear_135_cast_fp16")];
+            fp16 var_3613_to_fp16 = const()[name = string("op_3613_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_3614_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3613_to_fp16)[name = string("op_3614_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3614_cast_fp16)[name = string("input_803_cast_fp16")];
+            tensor<int32, [1]> input_805_axes_0 = const()[name = string("input_805_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304479936)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304482048)))];
+            tensor<fp16, [1, 28, 1024]> input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = string("input_805_cast_fp16")];
+            tensor<int32, [4]> cache_61_begin_0 = const()[name = string("cache_61_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_61_end_0 = const()[name = string("cache_61_end_0"), val = tensor<int32, [4]>([16, 1, 42, 1024])];
+            tensor<bool, [4]> cache_61_end_mask_0 = const()[name = string("cache_61_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_61_squeeze_mask_0 = const()[name = string("cache_61_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_61_cast_fp16 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_61_cast_fp16")];
+            tensor<int32, [4]> cache_63_begin_0 = const()[name = string("cache_63_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_63_end_0 = const()[name = string("cache_63_end_0"), val = tensor<int32, [4]>([16, 1, 1024, 8])];
+            tensor<bool, [4]> cache_63_end_mask_0 = const()[name = string("cache_63_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_63_squeeze_mask_0 = const()[name = string("cache_63_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_63_cast_fp16 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_63_cast_fp16")];
+            tensor<int32, [1]> input_807_axes_0 = const()[name = string("input_807_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304484160)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304486272)))];
+            tensor<fp16, [1, 28, 1024]> input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = string("input_807_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304488384))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307634176))))[name = string("encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_15_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307634368)))];
+            tensor<fp16, [1, 28, 4096]> linear_136_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized, x = input_807_cast_fp16)[name = string("linear_136_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = string("input_811_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(307642624))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310788416))))[name = string("encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310788608)))];
+            tensor<fp16, [1, 28, 1024]> linear_137_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("linear_137_cast_fp16")];
+            fp16 var_3650_to_fp16 = const()[name = string("op_3650_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_3651_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3650_to_fp16)[name = string("op_3651_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3651_cast_fp16)[name = string("input_817_cast_fp16")];
+            tensor<int32, [1]> key_31_axes_0 = const()[name = string("key_31_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310790720)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310792832)))];
+            tensor<fp16, [1, 28, 1024]> key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = string("key_31_cast_fp16")];
+            bool input_819_interleave_0 = const()[name = string("input_819_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_819_cast_fp16 = concat(axis = var_68, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = string("input_819_cast_fp16")];
+            tensor<int32, [3]> var_3673_begin_0 = const()[name = string("op_3673_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_3673_end_0 = const()[name = string("op_3673_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3673_end_mask_0 = const()[name = string("op_3673_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_3673_cast_fp16 = slice_by_index(begin = var_3673_begin_0, end = var_3673_end_0, end_mask = var_3673_end_mask_0, x = cache_61_cast_fp16)[name = string("op_3673_cast_fp16")];
+            bool var_3679_interleave_0 = const()[name = string("op_3679_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3679_cast_fp16 = concat(axis = var_68, interleave = var_3679_interleave_0, values = (var_3673_cast_fp16, key_31_cast_fp16))[name = string("op_3679_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310794944))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311581440))))[name = string("encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311581632)))];
+            tensor<fp16, [1, 28, 1024]> linear_138_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized, x = key_31_cast_fp16)[name = string("linear_138_cast_fp16")];
+            tensor<int32, [4]> var_3684 = const()[name = string("op_3684"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_91_cast_fp16 = reshape(shape = var_3684, x = linear_138_cast_fp16)[name = string("q_91_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311583744))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312370240))))[name = string("encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312370432)))];
+            tensor<fp16, [1, 70, 1024]> linear_139_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_139_cast_fp16")];
+            tensor<int32, [4]> var_3689 = const()[name = string("op_3689"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_61_cast_fp16 = reshape(shape = var_3689, x = linear_139_cast_fp16)[name = string("k_61_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312372544))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313159040))))[name = string("encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313159232)))];
+            tensor<fp16, [1, 70, 1024]> linear_140_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_140_cast_fp16")];
+            tensor<int32, [4]> var_3694 = const()[name = string("op_3694"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_31_cast_fp16 = reshape(shape = var_3694, x = linear_140_cast_fp16)[name = string("v_31_cast_fp16")];
+            tensor<int32, [4]> value_39_perm_0 = const()[name = string("value_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313161344)))];
+            tensor<fp16, [1, 28, 8, 128]> var_3707_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = string("op_3707_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313163456)))];
+            tensor<fp16, [1, 28, 8, 128]> var_3709_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = string("op_3709_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_31_perm_0 = const()[name = string("q_with_bias_v_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_397_transpose_x_0 = const()[name = string("x_397_transpose_x_0"), val = bool(false)];
+            bool x_397_transpose_y_0 = const()[name = string("x_397_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_3711_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313165568))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313307968))))[name = string("op_3711_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3709_cast_fp16)[name = string("transpose_227")];
+            tensor<fp16, [1, 8, 28, 139]> x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = op_3711_to_fp16_quantized)[name = string("x_397_cast_fp16")];
+            tensor<int32, [8]> x_399_pad_0 = const()[name = string("x_399_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_399_mode_0 = const()[name = string("x_399_mode_0"), val = string("constant")];
+            fp16 const_274_to_fp16 = const()[name = string("const_274_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_399_cast_fp16 = pad(constant_val = const_274_to_fp16, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397_cast_fp16)[name = string("x_399_cast_fp16")];
+            tensor<int32, [4]> var_3719 = const()[name = string("op_3719"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_401_cast_fp16 = reshape(shape = var_3719, x = x_399_cast_fp16)[name = string("x_401_cast_fp16")];
+            tensor<int32, [4]> var_3723_begin_0 = const()[name = string("op_3723_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3723_end_0 = const()[name = string("op_3723_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_3723_end_mask_0 = const()[name = string("op_3723_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_3723_cast_fp16 = slice_by_index(begin = var_3723_begin_0, end = var_3723_end_0, end_mask = var_3723_end_mask_0, x = x_401_cast_fp16)[name = string("op_3723_cast_fp16")];
+            tensor<int32, [4]> var_3724 = const()[name = string("op_3724"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_61_cast_fp16 = reshape(shape = var_3724, x = var_3723_cast_fp16)[name = string("matrix_bd_61_cast_fp16")];
+            bool matrix_ac_31_transpose_x_0 = const()[name = string("matrix_ac_31_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_31_transpose_y_0 = const()[name = string("matrix_ac_31_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_126_perm_0 = const()[name = string("transpose_126_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_127_perm_0 = const()[name = string("transpose_127_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_127 = transpose(perm = transpose_127_perm_0, x = k_61_cast_fp16)[name = string("transpose_225")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_126 = transpose(perm = transpose_126_perm_0, x = var_3707_cast_fp16)[name = string("transpose_226")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_31_cast_fp16 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_126, y = transpose_127)[name = string("matrix_ac_31_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_63_begin_0 = const()[name = string("matrix_bd_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_63_end_0 = const()[name = string("matrix_bd_63_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_63_end_mask_0 = const()[name = string("matrix_bd_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_63_cast_fp16 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61_cast_fp16)[name = string("matrix_bd_63_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_3733_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = string("op_3733_cast_fp16")];
+            fp16 _inversed_scores_61_y_0_to_fp16 = const()[name = string("_inversed_scores_61_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_61_cast_fp16 = mul(x = var_3733_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = string("_inversed_scores_61_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_63_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_11)[name = string("scores_63_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_3739_cast_fp16 = softmax(axis = var_59, x = scores_63_cast_fp16)[name = string("op_3739_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_821_cast_fp16 = select(a = var_44_to_fp16, b = var_3739_cast_fp16, cond = mask_11)[name = string("input_821_cast_fp16")];
+            bool x_403_transpose_x_0 = const()[name = string("x_403_transpose_x_0"), val = bool(false)];
+            bool x_403_transpose_y_0 = const()[name = string("x_403_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = v_31_cast_fp16)[name = string("transpose_224")];
+            tensor<fp16, [1, 8, 28, 128]> x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821_cast_fp16, y = value_39_cast_fp16)[name = string("x_403_cast_fp16")];
+            tensor<int32, [4]> var_3743_perm_0 = const()[name = string("op_3743_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3744 = const()[name = string("op_3744"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_3743_cast_fp16 = transpose(perm = var_3743_perm_0, x = x_403_cast_fp16)[name = string("transpose_223")];
+            tensor<fp16, [1, 28, 1024]> input_823_cast_fp16 = reshape(shape = var_3744, x = var_3743_cast_fp16)[name = string("input_823_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313308352))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314094848))))[name = string("encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314095040)))];
+            tensor<fp16, [1, 28, 1024]> linear_142_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized, x = input_823_cast_fp16)[name = string("linear_142_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = string("input_827_cast_fp16")];
+            tensor<int32, [1]> x_407_axes_0 = const()[name = string("x_407_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314097152)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314099264)))];
+            tensor<fp16, [1, 28, 1024]> x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = string("x_407_cast_fp16")];
+            tensor<int32, [3]> input_829_perm_0 = const()[name = string("input_829_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_831_pad_type_0 = const()[name = string("input_831_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_831_strides_0 = const()[name = string("input_831_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_831_pad_0 = const()[name = string("input_831_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_831_dilations_0 = const()[name = string("input_831_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_831_groups_0 = const()[name = string("input_831_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314101376))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316198592))))[name = string("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = string("transpose_222")];
+            tensor<fp16, [1, 2048, 28]> input_831_cast_fp16 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_829_cast_fp16)[name = string("input_831_cast_fp16")];
+            int32 x_409_split_num_splits_0 = const()[name = string("x_409_split_num_splits_0"), val = int32(2)];
+            int32 x_409_split_axis_0 = const()[name = string("x_409_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_409_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_409_split_cast_fp16_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831_cast_fp16)[name = string("x_409_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = string("x_409_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = string("x_409_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_833_cast_fp16 = select(a = var_44_to_fp16, b = x_409_cast_fp16, cond = var_575)[name = string("input_833_cast_fp16")];
+            bool new_x_63_interleave_0 = const()[name = string("new_x_63_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_63_cast_fp16 = concat(axis = var_59, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = string("new_x_63_cast_fp16")];
+            tensor<int32, [3]> var_3783_begin_0 = const()[name = string("op_3783_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_3783_end_0 = const()[name = string("op_3783_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_3783_end_mask_0 = const()[name = string("op_3783_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3783_cast_fp16 = slice_by_index(begin = var_3783_begin_0, end = var_3783_end_0, end_mask = var_3783_end_mask_0, x = new_x_63_cast_fp16)[name = string("op_3783_cast_fp16")];
+            string x_411_pad_type_0 = const()[name = string("x_411_pad_type_0"), val = string("valid")];
+            int32 x_411_groups_0 = const()[name = string("x_411_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_411_strides_0 = const()[name = string("x_411_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_411_pad_0 = const()[name = string("x_411_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_411_dilations_0 = const()[name = string("x_411_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316202752))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316212032))))[name = string("encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_411_cast_fp16 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_63_cast_fp16)[name = string("x_411_cast_fp16")];
+            tensor<int32, [3]> input_835_perm_0 = const()[name = string("input_835_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_413_axes_0 = const()[name = string("x_413_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316214144)))];
+            tensor<fp16, [1024]> encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316216256)))];
+            tensor<fp16, [1, 28, 1024]> input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = string("transpose_221")];
+            tensor<fp16, [1, 28, 1024]> x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = string("x_413_cast_fp16")];
+            tensor<int32, [3]> input_837_perm_0 = const()[name = string("input_837_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = string("transpose_220")];
+            tensor<fp16, [1, 1024, 28]> input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = string("input_839_cast_fp16")];
+            string x_415_pad_type_0 = const()[name = string("x_415_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_415_strides_0 = const()[name = string("x_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_415_pad_0 = const()[name = string("x_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_415_dilations_0 = const()[name = string("x_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_415_groups_0 = const()[name = string("x_415_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316218368))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317267008))))[name = string("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_839_cast_fp16)[name = string("x_415_cast_fp16")];
+            tensor<int32, [3]> input_841_perm_0 = const()[name = string("input_841_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = string("transpose_219")];
+            tensor<fp16, [1, 28, 1024]> input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = string("input_843_cast_fp16")];
+            tensor<int32, [1]> input_845_axes_0 = const()[name = string("input_845_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317269120)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317271232)))];
+            tensor<fp16, [1, 28, 1024]> input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = string("input_845_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317273344))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320419136))))[name = string("encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_15_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320419328)))];
+            tensor<fp16, [1, 28, 4096]> linear_143_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized, x = input_845_cast_fp16)[name = string("linear_143_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = string("input_849_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320427584))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323573376))))[name = string("encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323573568)))];
+            tensor<fp16, [1, 28, 1024]> linear_144_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized, x = input_849_cast_fp16)[name = string("linear_144_cast_fp16")];
+            fp16 var_3826_to_fp16 = const()[name = string("op_3826_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_3827_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3826_to_fp16)[name = string("op_3827_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3827_cast_fp16)[name = string("input_855_cast_fp16")];
+            tensor<int32, [1]> input_857_axes_0 = const()[name = string("input_857_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323575680)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323577792)))];
+            tensor<fp16, [1, 28, 1024]> input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = string("input_857_cast_fp16")];
+            tensor<int32, [4]> cache_65_begin_0 = const()[name = string("cache_65_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_65_end_0 = const()[name = string("cache_65_end_0"), val = tensor<int32, [4]>([17, 1, 42, 1024])];
+            tensor<bool, [4]> cache_65_end_mask_0 = const()[name = string("cache_65_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_65_squeeze_mask_0 = const()[name = string("cache_65_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_65_cast_fp16 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_65_cast_fp16")];
+            tensor<int32, [4]> cache_67_begin_0 = const()[name = string("cache_67_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_67_end_0 = const()[name = string("cache_67_end_0"), val = tensor<int32, [4]>([17, 1, 1024, 8])];
+            tensor<bool, [4]> cache_67_end_mask_0 = const()[name = string("cache_67_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_67_squeeze_mask_0 = const()[name = string("cache_67_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_67_cast_fp16 = slice_by_index(begin = cache_67_begin_0, end = cache_67_end_0, end_mask = cache_67_end_mask_0, squeeze_mask = cache_67_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_67_cast_fp16")];
+            tensor<int32, [1]> input_859_axes_0 = const()[name = string("input_859_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323579904)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323582016)))];
+            tensor<fp16, [1, 28, 1024]> input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = string("input_859_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323584128))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326729920))))[name = string("encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_16_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326730112)))];
+            tensor<fp16, [1, 28, 4096]> linear_145_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized, x = input_859_cast_fp16)[name = string("linear_145_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = string("input_863_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326738368))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329884160))))[name = string("encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329884352)))];
+            tensor<fp16, [1, 28, 1024]> linear_146_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized, x = input_863_cast_fp16)[name = string("linear_146_cast_fp16")];
+            fp16 var_3863_to_fp16 = const()[name = string("op_3863_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_3864_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3863_to_fp16)[name = string("op_3864_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3864_cast_fp16)[name = string("input_869_cast_fp16")];
+            tensor<int32, [1]> key_33_axes_0 = const()[name = string("key_33_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329886464)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329888576)))];
+            tensor<fp16, [1, 28, 1024]> key_33_cast_fp16 = layer_norm(axes = key_33_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = string("key_33_cast_fp16")];
+            bool input_871_interleave_0 = const()[name = string("input_871_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_871_cast_fp16 = concat(axis = var_68, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_33_cast_fp16))[name = string("input_871_cast_fp16")];
+            tensor<int32, [3]> var_3886_begin_0 = const()[name = string("op_3886_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_3886_end_0 = const()[name = string("op_3886_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3886_end_mask_0 = const()[name = string("op_3886_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_3886_cast_fp16 = slice_by_index(begin = var_3886_begin_0, end = var_3886_end_0, end_mask = var_3886_end_mask_0, x = cache_65_cast_fp16)[name = string("op_3886_cast_fp16")];
+            bool var_3892_interleave_0 = const()[name = string("op_3892_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3892_cast_fp16 = concat(axis = var_68, interleave = var_3892_interleave_0, values = (var_3886_cast_fp16, key_33_cast_fp16))[name = string("op_3892_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329890688))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330677184))))[name = string("encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330677376)))];
+            tensor<fp16, [1, 28, 1024]> linear_147_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized, x = key_33_cast_fp16)[name = string("linear_147_cast_fp16")];
+            tensor<int32, [4]> var_3897 = const()[name = string("op_3897"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_97_cast_fp16 = reshape(shape = var_3897, x = linear_147_cast_fp16)[name = string("q_97_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330679488))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331465984))))[name = string("encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331466176)))];
+            tensor<fp16, [1, 70, 1024]> linear_148_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_148_cast_fp16")];
+            tensor<int32, [4]> var_3902 = const()[name = string("op_3902"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_65_cast_fp16 = reshape(shape = var_3902, x = linear_148_cast_fp16)[name = string("k_65_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331468288))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332254784))))[name = string("encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332254976)))];
+            tensor<fp16, [1, 70, 1024]> linear_149_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_149_cast_fp16")];
+            tensor<int32, [4]> var_3907 = const()[name = string("op_3907"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_33_cast_fp16 = reshape(shape = var_3907, x = linear_149_cast_fp16)[name = string("v_33_cast_fp16")];
+            tensor<int32, [4]> value_41_perm_0 = const()[name = string("value_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332257088)))];
+            tensor<fp16, [1, 28, 8, 128]> var_3920_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = string("op_3920_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332259200)))];
+            tensor<fp16, [1, 28, 8, 128]> var_3922_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = string("op_3922_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_33_perm_0 = const()[name = string("q_with_bias_v_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_423_transpose_x_0 = const()[name = string("x_423_transpose_x_0"), val = bool(false)];
+            bool x_423_transpose_y_0 = const()[name = string("x_423_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_3924_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332261312))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332403712))))[name = string("op_3924_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_33_cast_fp16 = transpose(perm = q_with_bias_v_33_perm_0, x = var_3922_cast_fp16)[name = string("transpose_218")];
+            tensor<fp16, [1, 8, 28, 139]> x_423_cast_fp16 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_33_cast_fp16, y = op_3924_to_fp16_quantized)[name = string("x_423_cast_fp16")];
+            tensor<int32, [8]> x_425_pad_0 = const()[name = string("x_425_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_425_mode_0 = const()[name = string("x_425_mode_0"), val = string("constant")];
+            fp16 const_287_to_fp16 = const()[name = string("const_287_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_425_cast_fp16 = pad(constant_val = const_287_to_fp16, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423_cast_fp16)[name = string("x_425_cast_fp16")];
+            tensor<int32, [4]> var_3932 = const()[name = string("op_3932"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_427_cast_fp16 = reshape(shape = var_3932, x = x_425_cast_fp16)[name = string("x_427_cast_fp16")];
+            tensor<int32, [4]> var_3936_begin_0 = const()[name = string("op_3936_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3936_end_0 = const()[name = string("op_3936_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_3936_end_mask_0 = const()[name = string("op_3936_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_3936_cast_fp16 = slice_by_index(begin = var_3936_begin_0, end = var_3936_end_0, end_mask = var_3936_end_mask_0, x = x_427_cast_fp16)[name = string("op_3936_cast_fp16")];
+            tensor<int32, [4]> var_3937 = const()[name = string("op_3937"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_65_cast_fp16 = reshape(shape = var_3937, x = var_3936_cast_fp16)[name = string("matrix_bd_65_cast_fp16")];
+            bool matrix_ac_33_transpose_x_0 = const()[name = string("matrix_ac_33_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_33_transpose_y_0 = const()[name = string("matrix_ac_33_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_128_perm_0 = const()[name = string("transpose_128_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_129_perm_0 = const()[name = string("transpose_129_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_129 = transpose(perm = transpose_129_perm_0, x = k_65_cast_fp16)[name = string("transpose_216")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_128 = transpose(perm = transpose_128_perm_0, x = var_3920_cast_fp16)[name = string("transpose_217")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_33_cast_fp16 = matmul(transpose_x = matrix_ac_33_transpose_x_0, transpose_y = matrix_ac_33_transpose_y_0, x = transpose_128, y = transpose_129)[name = string("matrix_ac_33_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_67_begin_0 = const()[name = string("matrix_bd_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_67_end_0 = const()[name = string("matrix_bd_67_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_67_end_mask_0 = const()[name = string("matrix_bd_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_67_cast_fp16 = slice_by_index(begin = matrix_bd_67_begin_0, end = matrix_bd_67_end_0, end_mask = matrix_bd_67_end_mask_0, x = matrix_bd_65_cast_fp16)[name = string("matrix_bd_67_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_3946_cast_fp16 = add(x = matrix_ac_33_cast_fp16, y = matrix_bd_67_cast_fp16)[name = string("op_3946_cast_fp16")];
+            fp16 _inversed_scores_65_y_0_to_fp16 = const()[name = string("_inversed_scores_65_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_65_cast_fp16 = mul(x = var_3946_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = string("_inversed_scores_65_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_67_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_11)[name = string("scores_67_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_3952_cast_fp16 = softmax(axis = var_59, x = scores_67_cast_fp16)[name = string("op_3952_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_873_cast_fp16 = select(a = var_44_to_fp16, b = var_3952_cast_fp16, cond = mask_11)[name = string("input_873_cast_fp16")];
+            bool x_429_transpose_x_0 = const()[name = string("x_429_transpose_x_0"), val = bool(false)];
+            bool x_429_transpose_y_0 = const()[name = string("x_429_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = v_33_cast_fp16)[name = string("transpose_215")];
+            tensor<fp16, [1, 8, 28, 128]> x_429_cast_fp16 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873_cast_fp16, y = value_41_cast_fp16)[name = string("x_429_cast_fp16")];
+            tensor<int32, [4]> var_3956_perm_0 = const()[name = string("op_3956_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3957 = const()[name = string("op_3957"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_3956_cast_fp16 = transpose(perm = var_3956_perm_0, x = x_429_cast_fp16)[name = string("transpose_214")];
+            tensor<fp16, [1, 28, 1024]> input_875_cast_fp16 = reshape(shape = var_3957, x = var_3956_cast_fp16)[name = string("input_875_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332404096))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333190592))))[name = string("encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333190784)))];
+            tensor<fp16, [1, 28, 1024]> linear_151_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized, x = input_875_cast_fp16)[name = string("linear_151_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = string("input_879_cast_fp16")];
+            tensor<int32, [1]> x_433_axes_0 = const()[name = string("x_433_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333192896)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333195008)))];
+            tensor<fp16, [1, 28, 1024]> x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = string("x_433_cast_fp16")];
+            tensor<int32, [3]> input_881_perm_0 = const()[name = string("input_881_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_883_pad_type_0 = const()[name = string("input_883_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_883_strides_0 = const()[name = string("input_883_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_883_pad_0 = const()[name = string("input_883_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_883_dilations_0 = const()[name = string("input_883_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_883_groups_0 = const()[name = string("input_883_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333197120))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335294336))))[name = string("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = string("transpose_213")];
+            tensor<fp16, [1, 2048, 28]> input_883_cast_fp16 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_881_cast_fp16)[name = string("input_883_cast_fp16")];
+            int32 x_435_split_num_splits_0 = const()[name = string("x_435_split_num_splits_0"), val = int32(2)];
+            int32 x_435_split_axis_0 = const()[name = string("x_435_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_435_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_435_split_cast_fp16_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883_cast_fp16)[name = string("x_435_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = string("x_435_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = string("x_435_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_885_cast_fp16 = select(a = var_44_to_fp16, b = x_435_cast_fp16, cond = var_575)[name = string("input_885_cast_fp16")];
+            bool new_x_67_interleave_0 = const()[name = string("new_x_67_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_67_cast_fp16 = concat(axis = var_59, interleave = new_x_67_interleave_0, values = (cache_67_cast_fp16, input_885_cast_fp16))[name = string("new_x_67_cast_fp16")];
+            tensor<int32, [3]> var_3996_begin_0 = const()[name = string("op_3996_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_3996_end_0 = const()[name = string("op_3996_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_3996_end_mask_0 = const()[name = string("op_3996_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3996_cast_fp16 = slice_by_index(begin = var_3996_begin_0, end = var_3996_end_0, end_mask = var_3996_end_mask_0, x = new_x_67_cast_fp16)[name = string("op_3996_cast_fp16")];
+            string x_437_pad_type_0 = const()[name = string("x_437_pad_type_0"), val = string("valid")];
+            int32 x_437_groups_0 = const()[name = string("x_437_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_437_strides_0 = const()[name = string("x_437_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_437_pad_0 = const()[name = string("x_437_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_437_dilations_0 = const()[name = string("x_437_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335298496))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335307776))))[name = string("encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_437_cast_fp16 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_67_cast_fp16)[name = string("x_437_cast_fp16")];
+            tensor<int32, [3]> input_887_perm_0 = const()[name = string("input_887_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_439_axes_0 = const()[name = string("x_439_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335309888)))];
+            tensor<fp16, [1024]> encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335312000)))];
+            tensor<fp16, [1, 28, 1024]> input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = string("transpose_212")];
+            tensor<fp16, [1, 28, 1024]> x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = string("x_439_cast_fp16")];
+            tensor<int32, [3]> input_889_perm_0 = const()[name = string("input_889_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = string("transpose_211")];
+            tensor<fp16, [1, 1024, 28]> input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = string("input_891_cast_fp16")];
+            string x_441_pad_type_0 = const()[name = string("x_441_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_441_strides_0 = const()[name = string("x_441_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_441_pad_0 = const()[name = string("x_441_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_441_dilations_0 = const()[name = string("x_441_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_441_groups_0 = const()[name = string("x_441_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335314112))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336362752))))[name = string("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_441_cast_fp16 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_891_cast_fp16)[name = string("x_441_cast_fp16")];
+            tensor<int32, [3]> input_893_perm_0 = const()[name = string("input_893_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = string("transpose_210")];
+            tensor<fp16, [1, 28, 1024]> input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = string("input_895_cast_fp16")];
+            tensor<int32, [1]> input_897_axes_0 = const()[name = string("input_897_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336364864)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336366976)))];
+            tensor<fp16, [1, 28, 1024]> input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = string("input_897_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336369088))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339514880))))[name = string("encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_16_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339515072)))];
+            tensor<fp16, [1, 28, 4096]> linear_152_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized, x = input_897_cast_fp16)[name = string("linear_152_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = string("input_901_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339523328))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342669120))))[name = string("encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342669312)))];
+            tensor<fp16, [1, 28, 1024]> linear_153_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized, x = input_901_cast_fp16)[name = string("linear_153_cast_fp16")];
+            fp16 var_4039_to_fp16 = const()[name = string("op_4039_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_4040_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_4039_to_fp16)[name = string("op_4040_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_907_cast_fp16 = add(x = input_895_cast_fp16, y = var_4040_cast_fp16)[name = string("input_907_cast_fp16")];
+            tensor<int32, [1]> input_909_axes_0 = const()[name = string("input_909_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342671424)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342673536)))];
+            tensor<fp16, [1, 28, 1024]> input_909_cast_fp16 = layer_norm(axes = input_909_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_907_cast_fp16)[name = string("input_909_cast_fp16")];
+            tensor<int32, [4]> cache_69_begin_0 = const()[name = string("cache_69_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> cache_69_end_0 = const()[name = string("cache_69_end_0"), val = tensor<int32, [4]>([18, 1, 42, 1024])];
+            tensor<bool, [4]> cache_69_end_mask_0 = const()[name = string("cache_69_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_69_squeeze_mask_0 = const()[name = string("cache_69_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_69_cast_fp16 = slice_by_index(begin = cache_69_begin_0, end = cache_69_end_0, end_mask = cache_69_end_mask_0, squeeze_mask = cache_69_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_69_cast_fp16")];
+            tensor<int32, [4]> cache_71_begin_0 = const()[name = string("cache_71_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> cache_71_end_0 = const()[name = string("cache_71_end_0"), val = tensor<int32, [4]>([18, 1, 1024, 8])];
+            tensor<bool, [4]> cache_71_end_mask_0 = const()[name = string("cache_71_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_71_squeeze_mask_0 = const()[name = string("cache_71_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_71_cast_fp16 = slice_by_index(begin = cache_71_begin_0, end = cache_71_end_0, end_mask = cache_71_end_mask_0, squeeze_mask = cache_71_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_71_cast_fp16")];
+            tensor<int32, [1]> input_911_axes_0 = const()[name = string("input_911_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342675648)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342677760)))];
+            tensor<fp16, [1, 28, 1024]> input_911_cast_fp16 = layer_norm(axes = input_911_axes_0, beta = encoder_layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_feed_forward1_weight_to_fp16, x = input_909_cast_fp16)[name = string("input_911_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342679872))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345825664))))[name = string("encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_17_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345825856)))];
+            tensor<fp16, [1, 28, 4096]> linear_154_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized, x = input_911_cast_fp16)[name = string("linear_154_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_915_cast_fp16 = silu(x = linear_154_cast_fp16)[name = string("input_915_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345834112))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348979904))))[name = string("encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348980096)))];
+            tensor<fp16, [1, 28, 1024]> linear_155_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized, x = input_915_cast_fp16)[name = string("linear_155_cast_fp16")];
+            fp16 var_4076_to_fp16 = const()[name = string("op_4076_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_4077_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_4076_to_fp16)[name = string("op_4077_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_921_cast_fp16 = add(x = input_909_cast_fp16, y = var_4077_cast_fp16)[name = string("input_921_cast_fp16")];
+            tensor<int32, [1]> key_35_axes_0 = const()[name = string("key_35_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348982208)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348984320)))];
+            tensor<fp16, [1, 28, 1024]> key_35_cast_fp16 = layer_norm(axes = key_35_axes_0, beta = encoder_layers_17_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_self_att_weight_to_fp16, x = input_921_cast_fp16)[name = string("key_35_cast_fp16")];
+            bool input_923_interleave_0 = const()[name = string("input_923_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_923_cast_fp16 = concat(axis = var_68, interleave = input_923_interleave_0, values = (cache_69_cast_fp16, key_35_cast_fp16))[name = string("input_923_cast_fp16")];
+            tensor<int32, [3]> var_4099_begin_0 = const()[name = string("op_4099_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_4099_end_0 = const()[name = string("op_4099_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4099_end_mask_0 = const()[name = string("op_4099_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_4099_cast_fp16 = slice_by_index(begin = var_4099_begin_0, end = var_4099_end_0, end_mask = var_4099_end_mask_0, x = cache_69_cast_fp16)[name = string("op_4099_cast_fp16")];
+            bool var_4105_interleave_0 = const()[name = string("op_4105_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4105_cast_fp16 = concat(axis = var_68, interleave = var_4105_interleave_0, values = (var_4099_cast_fp16, key_35_cast_fp16))[name = string("op_4105_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348986432))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349772928))))[name = string("encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349773120)))];
+            tensor<fp16, [1, 28, 1024]> linear_156_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized, x = key_35_cast_fp16)[name = string("linear_156_cast_fp16")];
+            tensor<int32, [4]> var_4110 = const()[name = string("op_4110"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_103_cast_fp16 = reshape(shape = var_4110, x = linear_156_cast_fp16)[name = string("q_103_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349775232))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350561728))))[name = string("encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350561920)))];
+            tensor<fp16, [1, 70, 1024]> linear_157_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_157_cast_fp16")];
+            tensor<int32, [4]> var_4115 = const()[name = string("op_4115"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_69_cast_fp16 = reshape(shape = var_4115, x = linear_157_cast_fp16)[name = string("k_69_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350564032))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351350528))))[name = string("encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351350720)))];
+            tensor<fp16, [1, 70, 1024]> linear_158_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_158_cast_fp16")];
+            tensor<int32, [4]> var_4120 = const()[name = string("op_4120"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_35_cast_fp16 = reshape(shape = var_4120, x = linear_158_cast_fp16)[name = string("v_35_cast_fp16")];
+            tensor<int32, [4]> value_43_perm_0 = const()[name = string("value_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_17_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351352832)))];
+            tensor<fp16, [1, 28, 8, 128]> var_4133_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_u_to_fp16)[name = string("op_4133_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_17_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351354944)))];
+            tensor<fp16, [1, 28, 8, 128]> var_4135_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_v_to_fp16)[name = string("op_4135_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_35_perm_0 = const()[name = string("q_with_bias_v_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_449_transpose_x_0 = const()[name = string("x_449_transpose_x_0"), val = bool(false)];
+            bool x_449_transpose_y_0 = const()[name = string("x_449_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_4137_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351357056))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351499456))))[name = string("op_4137_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_35_cast_fp16 = transpose(perm = q_with_bias_v_35_perm_0, x = var_4135_cast_fp16)[name = string("transpose_209")];
+            tensor<fp16, [1, 8, 28, 139]> x_449_cast_fp16 = matmul(transpose_x = x_449_transpose_x_0, transpose_y = x_449_transpose_y_0, x = q_with_bias_v_35_cast_fp16, y = op_4137_to_fp16_quantized)[name = string("x_449_cast_fp16")];
+            tensor<int32, [8]> x_451_pad_0 = const()[name = string("x_451_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_451_mode_0 = const()[name = string("x_451_mode_0"), val = string("constant")];
+            fp16 const_300_to_fp16 = const()[name = string("const_300_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_451_cast_fp16 = pad(constant_val = const_300_to_fp16, mode = x_451_mode_0, pad = x_451_pad_0, x = x_449_cast_fp16)[name = string("x_451_cast_fp16")];
+            tensor<int32, [4]> var_4145 = const()[name = string("op_4145"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_453_cast_fp16 = reshape(shape = var_4145, x = x_451_cast_fp16)[name = string("x_453_cast_fp16")];
+            tensor<int32, [4]> var_4149_begin_0 = const()[name = string("op_4149_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4149_end_0 = const()[name = string("op_4149_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_4149_end_mask_0 = const()[name = string("op_4149_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_4149_cast_fp16 = slice_by_index(begin = var_4149_begin_0, end = var_4149_end_0, end_mask = var_4149_end_mask_0, x = x_453_cast_fp16)[name = string("op_4149_cast_fp16")];
+            tensor<int32, [4]> var_4150 = const()[name = string("op_4150"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_69_cast_fp16 = reshape(shape = var_4150, x = var_4149_cast_fp16)[name = string("matrix_bd_69_cast_fp16")];
+            bool matrix_ac_35_transpose_x_0 = const()[name = string("matrix_ac_35_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_35_transpose_y_0 = const()[name = string("matrix_ac_35_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_130_perm_0 = const()[name = string("transpose_130_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_131_perm_0 = const()[name = string("transpose_131_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_131 = transpose(perm = transpose_131_perm_0, x = k_69_cast_fp16)[name = string("transpose_207")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_130 = transpose(perm = transpose_130_perm_0, x = var_4133_cast_fp16)[name = string("transpose_208")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_35_cast_fp16 = matmul(transpose_x = matrix_ac_35_transpose_x_0, transpose_y = matrix_ac_35_transpose_y_0, x = transpose_130, y = transpose_131)[name = string("matrix_ac_35_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_71_begin_0 = const()[name = string("matrix_bd_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_71_end_0 = const()[name = string("matrix_bd_71_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_71_end_mask_0 = const()[name = string("matrix_bd_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_71_cast_fp16 = slice_by_index(begin = matrix_bd_71_begin_0, end = matrix_bd_71_end_0, end_mask = matrix_bd_71_end_mask_0, x = matrix_bd_69_cast_fp16)[name = string("matrix_bd_71_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_4159_cast_fp16 = add(x = matrix_ac_35_cast_fp16, y = matrix_bd_71_cast_fp16)[name = string("op_4159_cast_fp16")];
+            fp16 _inversed_scores_69_y_0_to_fp16 = const()[name = string("_inversed_scores_69_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_69_cast_fp16 = mul(x = var_4159_cast_fp16, y = _inversed_scores_69_y_0_to_fp16)[name = string("_inversed_scores_69_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_71_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_69_cast_fp16, cond = mask_11)[name = string("scores_71_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_4165_cast_fp16 = softmax(axis = var_59, x = scores_71_cast_fp16)[name = string("op_4165_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_925_cast_fp16 = select(a = var_44_to_fp16, b = var_4165_cast_fp16, cond = mask_11)[name = string("input_925_cast_fp16")];
+            bool x_455_transpose_x_0 = const()[name = string("x_455_transpose_x_0"), val = bool(false)];
+            bool x_455_transpose_y_0 = const()[name = string("x_455_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = v_35_cast_fp16)[name = string("transpose_206")];
+            tensor<fp16, [1, 8, 28, 128]> x_455_cast_fp16 = matmul(transpose_x = x_455_transpose_x_0, transpose_y = x_455_transpose_y_0, x = input_925_cast_fp16, y = value_43_cast_fp16)[name = string("x_455_cast_fp16")];
+            tensor<int32, [4]> var_4169_perm_0 = const()[name = string("op_4169_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4170 = const()[name = string("op_4170"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_4169_cast_fp16 = transpose(perm = var_4169_perm_0, x = x_455_cast_fp16)[name = string("transpose_205")];
+            tensor<fp16, [1, 28, 1024]> input_927_cast_fp16 = reshape(shape = var_4170, x = var_4169_cast_fp16)[name = string("input_927_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351499840))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352286336))))[name = string("encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352286528)))];
+            tensor<fp16, [1, 28, 1024]> linear_160_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized, x = input_927_cast_fp16)[name = string("linear_160_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_931_cast_fp16 = add(x = input_921_cast_fp16, y = linear_160_cast_fp16)[name = string("input_931_cast_fp16")];
+            tensor<int32, [1]> x_459_axes_0 = const()[name = string("x_459_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352288640)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352290752)))];
+            tensor<fp16, [1, 28, 1024]> x_459_cast_fp16 = layer_norm(axes = x_459_axes_0, beta = encoder_layers_17_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_conv_weight_to_fp16, x = input_931_cast_fp16)[name = string("x_459_cast_fp16")];
+            tensor<int32, [3]> input_933_perm_0 = const()[name = string("input_933_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_935_pad_type_0 = const()[name = string("input_935_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_935_strides_0 = const()[name = string("input_935_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_935_pad_0 = const()[name = string("input_935_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_935_dilations_0 = const()[name = string("input_935_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_935_groups_0 = const()[name = string("input_935_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352292864))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354390080))))[name = string("encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_933_cast_fp16 = transpose(perm = input_933_perm_0, x = x_459_cast_fp16)[name = string("transpose_204")];
+            tensor<fp16, [1, 2048, 28]> input_935_cast_fp16 = conv(dilations = input_935_dilations_0, groups = input_935_groups_0, pad = input_935_pad_0, pad_type = input_935_pad_type_0, strides = input_935_strides_0, weight = encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_933_cast_fp16)[name = string("input_935_cast_fp16")];
+            int32 x_461_split_num_splits_0 = const()[name = string("x_461_split_num_splits_0"), val = int32(2)];
+            int32 x_461_split_axis_0 = const()[name = string("x_461_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_461_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_461_split_cast_fp16_1 = split(axis = x_461_split_axis_0, num_splits = x_461_split_num_splits_0, x = input_935_cast_fp16)[name = string("x_461_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_461_split_1_sigmoid_cast_fp16 = sigmoid(x = x_461_split_cast_fp16_1)[name = string("x_461_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_461_cast_fp16 = mul(x = x_461_split_cast_fp16_0, y = x_461_split_1_sigmoid_cast_fp16)[name = string("x_461_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_937_cast_fp16 = select(a = var_44_to_fp16, b = x_461_cast_fp16, cond = var_575)[name = string("input_937_cast_fp16")];
+            bool new_x_71_interleave_0 = const()[name = string("new_x_71_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_71_cast_fp16 = concat(axis = var_59, interleave = new_x_71_interleave_0, values = (cache_71_cast_fp16, input_937_cast_fp16))[name = string("new_x_71_cast_fp16")];
+            tensor<int32, [3]> var_4209_begin_0 = const()[name = string("op_4209_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_4209_end_0 = const()[name = string("op_4209_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_4209_end_mask_0 = const()[name = string("op_4209_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4209_cast_fp16 = slice_by_index(begin = var_4209_begin_0, end = var_4209_end_0, end_mask = var_4209_end_mask_0, x = new_x_71_cast_fp16)[name = string("op_4209_cast_fp16")];
+            string x_463_pad_type_0 = const()[name = string("x_463_pad_type_0"), val = string("valid")];
+            int32 x_463_groups_0 = const()[name = string("x_463_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_463_strides_0 = const()[name = string("x_463_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_463_pad_0 = const()[name = string("x_463_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_463_dilations_0 = const()[name = string("x_463_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354394240))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354403520))))[name = string("encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_463_cast_fp16 = conv(dilations = x_463_dilations_0, groups = x_463_groups_0, pad = x_463_pad_0, pad_type = x_463_pad_type_0, strides = x_463_strides_0, weight = encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_71_cast_fp16)[name = string("x_463_cast_fp16")];
+            tensor<int32, [3]> input_939_perm_0 = const()[name = string("input_939_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_465_axes_0 = const()[name = string("x_465_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354405632)))];
+            tensor<fp16, [1024]> encoder_layers_17_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354407744)))];
+            tensor<fp16, [1, 28, 1024]> input_939_cast_fp16 = transpose(perm = input_939_perm_0, x = x_463_cast_fp16)[name = string("transpose_203")];
+            tensor<fp16, [1, 28, 1024]> x_465_cast_fp16 = layer_norm(axes = x_465_axes_0, beta = encoder_layers_17_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_conv_batch_norm_weight_to_fp16, x = input_939_cast_fp16)[name = string("x_465_cast_fp16")];
+            tensor<int32, [3]> input_941_perm_0 = const()[name = string("input_941_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_941_cast_fp16 = transpose(perm = input_941_perm_0, x = x_465_cast_fp16)[name = string("transpose_202")];
+            tensor<fp16, [1, 1024, 28]> input_943_cast_fp16 = silu(x = input_941_cast_fp16)[name = string("input_943_cast_fp16")];
+            string x_467_pad_type_0 = const()[name = string("x_467_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_467_strides_0 = const()[name = string("x_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_467_pad_0 = const()[name = string("x_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_467_dilations_0 = const()[name = string("x_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_467_groups_0 = const()[name = string("x_467_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354409856))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355458496))))[name = string("encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_467_cast_fp16 = conv(dilations = x_467_dilations_0, groups = x_467_groups_0, pad = x_467_pad_0, pad_type = x_467_pad_type_0, strides = x_467_strides_0, weight = encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_943_cast_fp16)[name = string("x_467_cast_fp16")];
+            tensor<int32, [3]> input_945_perm_0 = const()[name = string("input_945_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_945_cast_fp16 = transpose(perm = input_945_perm_0, x = x_467_cast_fp16)[name = string("transpose_201")];
+            tensor<fp16, [1, 28, 1024]> input_947_cast_fp16 = add(x = input_931_cast_fp16, y = input_945_cast_fp16)[name = string("input_947_cast_fp16")];
+            tensor<int32, [1]> input_949_axes_0 = const()[name = string("input_949_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355460608)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355462720)))];
+            tensor<fp16, [1, 28, 1024]> input_949_cast_fp16 = layer_norm(axes = input_949_axes_0, beta = encoder_layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_feed_forward2_weight_to_fp16, x = input_947_cast_fp16)[name = string("input_949_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355464832))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358610624))))[name = string("encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_17_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358610816)))];
+            tensor<fp16, [1, 28, 4096]> linear_161_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("linear_161_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_953_cast_fp16 = silu(x = linear_161_cast_fp16)[name = string("input_953_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358619072))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361764864))))[name = string("encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361765056)))];
+            tensor<fp16, [1, 28, 1024]> linear_162_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized, x = input_953_cast_fp16)[name = string("linear_162_cast_fp16")];
+            fp16 var_4252_to_fp16 = const()[name = string("op_4252_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_4253_cast_fp16 = mul(x = linear_162_cast_fp16, y = var_4252_to_fp16)[name = string("op_4253_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_959_cast_fp16 = add(x = input_947_cast_fp16, y = var_4253_cast_fp16)[name = string("input_959_cast_fp16")];
+            tensor<int32, [1]> input_961_axes_0 = const()[name = string("input_961_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361767168)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361769280)))];
+            tensor<fp16, [1, 28, 1024]> input_961_cast_fp16 = layer_norm(axes = input_961_axes_0, beta = encoder_layers_17_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_out_weight_to_fp16, x = input_959_cast_fp16)[name = string("input_961_cast_fp16")];
+            tensor<int32, [4]> cache_73_begin_0 = const()[name = string("cache_73_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> cache_73_end_0 = const()[name = string("cache_73_end_0"), val = tensor<int32, [4]>([19, 1, 42, 1024])];
+            tensor<bool, [4]> cache_73_end_mask_0 = const()[name = string("cache_73_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_73_squeeze_mask_0 = const()[name = string("cache_73_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_73_cast_fp16 = slice_by_index(begin = cache_73_begin_0, end = cache_73_end_0, end_mask = cache_73_end_mask_0, squeeze_mask = cache_73_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_73_cast_fp16")];
+            tensor<int32, [4]> cache_75_begin_0 = const()[name = string("cache_75_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> cache_75_end_0 = const()[name = string("cache_75_end_0"), val = tensor<int32, [4]>([19, 1, 1024, 8])];
+            tensor<bool, [4]> cache_75_end_mask_0 = const()[name = string("cache_75_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_75_squeeze_mask_0 = const()[name = string("cache_75_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_75_cast_fp16 = slice_by_index(begin = cache_75_begin_0, end = cache_75_end_0, end_mask = cache_75_end_mask_0, squeeze_mask = cache_75_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_75_cast_fp16")];
+            tensor<int32, [1]> input_963_axes_0 = const()[name = string("input_963_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361771392)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361773504)))];
+            tensor<fp16, [1, 28, 1024]> input_963_cast_fp16 = layer_norm(axes = input_963_axes_0, beta = encoder_layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_feed_forward1_weight_to_fp16, x = input_961_cast_fp16)[name = string("input_963_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361775616))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364921408))))[name = string("encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_18_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364921600)))];
+            tensor<fp16, [1, 28, 4096]> linear_163_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized, x = input_963_cast_fp16)[name = string("linear_163_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_967_cast_fp16 = silu(x = linear_163_cast_fp16)[name = string("input_967_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364929856))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368075648))))[name = string("encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368075840)))];
+            tensor<fp16, [1, 28, 1024]> linear_164_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized, x = input_967_cast_fp16)[name = string("linear_164_cast_fp16")];
+            fp16 var_4289_to_fp16 = const()[name = string("op_4289_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_4290_cast_fp16 = mul(x = linear_164_cast_fp16, y = var_4289_to_fp16)[name = string("op_4290_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_973_cast_fp16 = add(x = input_961_cast_fp16, y = var_4290_cast_fp16)[name = string("input_973_cast_fp16")];
+            tensor<int32, [1]> key_37_axes_0 = const()[name = string("key_37_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368077952)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368080064)))];
+            tensor<fp16, [1, 28, 1024]> key_37_cast_fp16 = layer_norm(axes = key_37_axes_0, beta = encoder_layers_18_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_self_att_weight_to_fp16, x = input_973_cast_fp16)[name = string("key_37_cast_fp16")];
+            bool input_975_interleave_0 = const()[name = string("input_975_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_975_cast_fp16 = concat(axis = var_68, interleave = input_975_interleave_0, values = (cache_73_cast_fp16, key_37_cast_fp16))[name = string("input_975_cast_fp16")];
+            tensor<int32, [3]> var_4312_begin_0 = const()[name = string("op_4312_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_4312_end_0 = const()[name = string("op_4312_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4312_end_mask_0 = const()[name = string("op_4312_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_4312_cast_fp16 = slice_by_index(begin = var_4312_begin_0, end = var_4312_end_0, end_mask = var_4312_end_mask_0, x = cache_73_cast_fp16)[name = string("op_4312_cast_fp16")];
+            bool var_4318_interleave_0 = const()[name = string("op_4318_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4318_cast_fp16 = concat(axis = var_68, interleave = var_4318_interleave_0, values = (var_4312_cast_fp16, key_37_cast_fp16))[name = string("op_4318_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368082176))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368868672))))[name = string("encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368868864)))];
+            tensor<fp16, [1, 28, 1024]> linear_165_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized, x = key_37_cast_fp16)[name = string("linear_165_cast_fp16")];
+            tensor<int32, [4]> var_4323 = const()[name = string("op_4323"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_109_cast_fp16 = reshape(shape = var_4323, x = linear_165_cast_fp16)[name = string("q_109_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368870976))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369657472))))[name = string("encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369657664)))];
+            tensor<fp16, [1, 70, 1024]> linear_166_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_166_cast_fp16")];
+            tensor<int32, [4]> var_4328 = const()[name = string("op_4328"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_73_cast_fp16 = reshape(shape = var_4328, x = linear_166_cast_fp16)[name = string("k_73_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369659776))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370446272))))[name = string("encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370446464)))];
+            tensor<fp16, [1, 70, 1024]> linear_167_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_167_cast_fp16")];
+            tensor<int32, [4]> var_4333 = const()[name = string("op_4333"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_37_cast_fp16 = reshape(shape = var_4333, x = linear_167_cast_fp16)[name = string("v_37_cast_fp16")];
+            tensor<int32, [4]> value_45_perm_0 = const()[name = string("value_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_18_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370448576)))];
+            tensor<fp16, [1, 28, 8, 128]> var_4346_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_u_to_fp16)[name = string("op_4346_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_18_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370450688)))];
+            tensor<fp16, [1, 28, 8, 128]> var_4348_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_v_to_fp16)[name = string("op_4348_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_37_perm_0 = const()[name = string("q_with_bias_v_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_475_transpose_x_0 = const()[name = string("x_475_transpose_x_0"), val = bool(false)];
+            bool x_475_transpose_y_0 = const()[name = string("x_475_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_4350_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370452800))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370595200))))[name = string("op_4350_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_37_cast_fp16 = transpose(perm = q_with_bias_v_37_perm_0, x = var_4348_cast_fp16)[name = string("transpose_200")];
+            tensor<fp16, [1, 8, 28, 139]> x_475_cast_fp16 = matmul(transpose_x = x_475_transpose_x_0, transpose_y = x_475_transpose_y_0, x = q_with_bias_v_37_cast_fp16, y = op_4350_to_fp16_quantized)[name = string("x_475_cast_fp16")];
+            tensor<int32, [8]> x_477_pad_0 = const()[name = string("x_477_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_477_mode_0 = const()[name = string("x_477_mode_0"), val = string("constant")];
+            fp16 const_313_to_fp16 = const()[name = string("const_313_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_477_cast_fp16 = pad(constant_val = const_313_to_fp16, mode = x_477_mode_0, pad = x_477_pad_0, x = x_475_cast_fp16)[name = string("x_477_cast_fp16")];
+            tensor<int32, [4]> var_4358 = const()[name = string("op_4358"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_479_cast_fp16 = reshape(shape = var_4358, x = x_477_cast_fp16)[name = string("x_479_cast_fp16")];
+            tensor<int32, [4]> var_4362_begin_0 = const()[name = string("op_4362_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4362_end_0 = const()[name = string("op_4362_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_4362_end_mask_0 = const()[name = string("op_4362_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_4362_cast_fp16 = slice_by_index(begin = var_4362_begin_0, end = var_4362_end_0, end_mask = var_4362_end_mask_0, x = x_479_cast_fp16)[name = string("op_4362_cast_fp16")];
+            tensor<int32, [4]> var_4363 = const()[name = string("op_4363"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_73_cast_fp16 = reshape(shape = var_4363, x = var_4362_cast_fp16)[name = string("matrix_bd_73_cast_fp16")];
+            bool matrix_ac_37_transpose_x_0 = const()[name = string("matrix_ac_37_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_37_transpose_y_0 = const()[name = string("matrix_ac_37_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_132_perm_0 = const()[name = string("transpose_132_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_133_perm_0 = const()[name = string("transpose_133_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_133 = transpose(perm = transpose_133_perm_0, x = k_73_cast_fp16)[name = string("transpose_198")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_132 = transpose(perm = transpose_132_perm_0, x = var_4346_cast_fp16)[name = string("transpose_199")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_37_cast_fp16 = matmul(transpose_x = matrix_ac_37_transpose_x_0, transpose_y = matrix_ac_37_transpose_y_0, x = transpose_132, y = transpose_133)[name = string("matrix_ac_37_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_75_begin_0 = const()[name = string("matrix_bd_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_75_end_0 = const()[name = string("matrix_bd_75_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_75_end_mask_0 = const()[name = string("matrix_bd_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_75_cast_fp16 = slice_by_index(begin = matrix_bd_75_begin_0, end = matrix_bd_75_end_0, end_mask = matrix_bd_75_end_mask_0, x = matrix_bd_73_cast_fp16)[name = string("matrix_bd_75_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_4372_cast_fp16 = add(x = matrix_ac_37_cast_fp16, y = matrix_bd_75_cast_fp16)[name = string("op_4372_cast_fp16")];
+            fp16 _inversed_scores_73_y_0_to_fp16 = const()[name = string("_inversed_scores_73_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_73_cast_fp16 = mul(x = var_4372_cast_fp16, y = _inversed_scores_73_y_0_to_fp16)[name = string("_inversed_scores_73_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_75_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_73_cast_fp16, cond = mask_11)[name = string("scores_75_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_4378_cast_fp16 = softmax(axis = var_59, x = scores_75_cast_fp16)[name = string("op_4378_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_977_cast_fp16 = select(a = var_44_to_fp16, b = var_4378_cast_fp16, cond = mask_11)[name = string("input_977_cast_fp16")];
+            bool x_481_transpose_x_0 = const()[name = string("x_481_transpose_x_0"), val = bool(false)];
+            bool x_481_transpose_y_0 = const()[name = string("x_481_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = v_37_cast_fp16)[name = string("transpose_197")];
+            tensor<fp16, [1, 8, 28, 128]> x_481_cast_fp16 = matmul(transpose_x = x_481_transpose_x_0, transpose_y = x_481_transpose_y_0, x = input_977_cast_fp16, y = value_45_cast_fp16)[name = string("x_481_cast_fp16")];
+            tensor<int32, [4]> var_4382_perm_0 = const()[name = string("op_4382_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4383 = const()[name = string("op_4383"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_4382_cast_fp16 = transpose(perm = var_4382_perm_0, x = x_481_cast_fp16)[name = string("transpose_196")];
+            tensor<fp16, [1, 28, 1024]> input_979_cast_fp16 = reshape(shape = var_4383, x = var_4382_cast_fp16)[name = string("input_979_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370595584))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371644224))))[name = string("encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371646336)))];
+            tensor<fp16, [1, 28, 1024]> linear_169_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized, x = input_979_cast_fp16)[name = string("linear_169_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_983_cast_fp16 = add(x = input_973_cast_fp16, y = linear_169_cast_fp16)[name = string("input_983_cast_fp16")];
+            tensor<int32, [1]> x_485_axes_0 = const()[name = string("x_485_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371648448)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371650560)))];
+            tensor<fp16, [1, 28, 1024]> x_485_cast_fp16 = layer_norm(axes = x_485_axes_0, beta = encoder_layers_18_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_conv_weight_to_fp16, x = input_983_cast_fp16)[name = string("x_485_cast_fp16")];
+            tensor<int32, [3]> input_985_perm_0 = const()[name = string("input_985_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_987_pad_type_0 = const()[name = string("input_987_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_987_strides_0 = const()[name = string("input_987_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_987_pad_0 = const()[name = string("input_987_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_987_dilations_0 = const()[name = string("input_987_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_987_groups_0 = const()[name = string("input_987_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371652672))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373749888))))[name = string("encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_985_cast_fp16 = transpose(perm = input_985_perm_0, x = x_485_cast_fp16)[name = string("transpose_195")];
+            tensor<fp16, [1, 2048, 28]> input_987_cast_fp16 = conv(dilations = input_987_dilations_0, groups = input_987_groups_0, pad = input_987_pad_0, pad_type = input_987_pad_type_0, strides = input_987_strides_0, weight = encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_985_cast_fp16)[name = string("input_987_cast_fp16")];
+            int32 x_487_split_num_splits_0 = const()[name = string("x_487_split_num_splits_0"), val = int32(2)];
+            int32 x_487_split_axis_0 = const()[name = string("x_487_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_487_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_487_split_cast_fp16_1 = split(axis = x_487_split_axis_0, num_splits = x_487_split_num_splits_0, x = input_987_cast_fp16)[name = string("x_487_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_487_split_1_sigmoid_cast_fp16 = sigmoid(x = x_487_split_cast_fp16_1)[name = string("x_487_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_487_cast_fp16 = mul(x = x_487_split_cast_fp16_0, y = x_487_split_1_sigmoid_cast_fp16)[name = string("x_487_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_989_cast_fp16 = select(a = var_44_to_fp16, b = x_487_cast_fp16, cond = var_575)[name = string("input_989_cast_fp16")];
+            bool new_x_75_interleave_0 = const()[name = string("new_x_75_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_75_cast_fp16 = concat(axis = var_59, interleave = new_x_75_interleave_0, values = (cache_75_cast_fp16, input_989_cast_fp16))[name = string("new_x_75_cast_fp16")];
+            tensor<int32, [3]> var_4422_begin_0 = const()[name = string("op_4422_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_4422_end_0 = const()[name = string("op_4422_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_4422_end_mask_0 = const()[name = string("op_4422_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4422_cast_fp16 = slice_by_index(begin = var_4422_begin_0, end = var_4422_end_0, end_mask = var_4422_end_mask_0, x = new_x_75_cast_fp16)[name = string("op_4422_cast_fp16")];
+            string x_489_pad_type_0 = const()[name = string("x_489_pad_type_0"), val = string("valid")];
+            int32 x_489_groups_0 = const()[name = string("x_489_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_489_strides_0 = const()[name = string("x_489_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_489_pad_0 = const()[name = string("x_489_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_489_dilations_0 = const()[name = string("x_489_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373754048))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373763328))))[name = string("encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_489_cast_fp16 = conv(dilations = x_489_dilations_0, groups = x_489_groups_0, pad = x_489_pad_0, pad_type = x_489_pad_type_0, strides = x_489_strides_0, weight = encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_75_cast_fp16)[name = string("x_489_cast_fp16")];
+            tensor<int32, [3]> input_991_perm_0 = const()[name = string("input_991_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_491_axes_0 = const()[name = string("x_491_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373765440)))];
+            tensor<fp16, [1024]> encoder_layers_18_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373767552)))];
+            tensor<fp16, [1, 28, 1024]> input_991_cast_fp16 = transpose(perm = input_991_perm_0, x = x_489_cast_fp16)[name = string("transpose_194")];
+            tensor<fp16, [1, 28, 1024]> x_491_cast_fp16 = layer_norm(axes = x_491_axes_0, beta = encoder_layers_18_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_conv_batch_norm_weight_to_fp16, x = input_991_cast_fp16)[name = string("x_491_cast_fp16")];
+            tensor<int32, [3]> input_993_perm_0 = const()[name = string("input_993_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_993_cast_fp16 = transpose(perm = input_993_perm_0, x = x_491_cast_fp16)[name = string("transpose_193")];
+            tensor<fp16, [1, 1024, 28]> input_995_cast_fp16 = silu(x = input_993_cast_fp16)[name = string("input_995_cast_fp16")];
+            string x_493_pad_type_0 = const()[name = string("x_493_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_493_strides_0 = const()[name = string("x_493_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_493_pad_0 = const()[name = string("x_493_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_493_dilations_0 = const()[name = string("x_493_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_493_groups_0 = const()[name = string("x_493_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373769664))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374818304))))[name = string("encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_493_cast_fp16 = conv(dilations = x_493_dilations_0, groups = x_493_groups_0, pad = x_493_pad_0, pad_type = x_493_pad_type_0, strides = x_493_strides_0, weight = encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_995_cast_fp16)[name = string("x_493_cast_fp16")];
+            tensor<int32, [3]> input_997_perm_0 = const()[name = string("input_997_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_997_cast_fp16 = transpose(perm = input_997_perm_0, x = x_493_cast_fp16)[name = string("transpose_192")];
+            tensor<fp16, [1, 28, 1024]> input_999_cast_fp16 = add(x = input_983_cast_fp16, y = input_997_cast_fp16)[name = string("input_999_cast_fp16")];
+            tensor<int32, [1]> input_1001_axes_0 = const()[name = string("input_1001_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374820416)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374822528)))];
+            tensor<fp16, [1, 28, 1024]> input_1001_cast_fp16 = layer_norm(axes = input_1001_axes_0, beta = encoder_layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_feed_forward2_weight_to_fp16, x = input_999_cast_fp16)[name = string("input_1001_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374824640))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379019008))))[name = string("encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_18_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379027264)))];
+            tensor<fp16, [1, 28, 4096]> linear_170_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1001_cast_fp16)[name = string("linear_170_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_1005_cast_fp16 = silu(x = linear_170_cast_fp16)[name = string("input_1005_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379035520))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383229888))))[name = string("encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_18_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383232000)))];
+            tensor<fp16, [1, 28, 1024]> linear_171_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1005_cast_fp16)[name = string("linear_171_cast_fp16")];
+            fp16 var_4465_to_fp16 = const()[name = string("op_4465_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_4466_cast_fp16 = mul(x = linear_171_cast_fp16, y = var_4465_to_fp16)[name = string("op_4466_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_1011_cast_fp16 = add(x = input_999_cast_fp16, y = var_4466_cast_fp16)[name = string("input_1011_cast_fp16")];
+            tensor<int32, [1]> input_1013_axes_0 = const()[name = string("input_1013_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383234112)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383236224)))];
+            tensor<fp16, [1, 28, 1024]> input_1013_cast_fp16 = layer_norm(axes = input_1013_axes_0, beta = encoder_layers_18_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_out_weight_to_fp16, x = input_1011_cast_fp16)[name = string("input_1013_cast_fp16")];
+            tensor<int32, [4]> cache_77_begin_0 = const()[name = string("cache_77_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
+            tensor<int32, [4]> cache_77_end_0 = const()[name = string("cache_77_end_0"), val = tensor<int32, [4]>([20, 1, 42, 1024])];
+            tensor<bool, [4]> cache_77_end_mask_0 = const()[name = string("cache_77_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_77_squeeze_mask_0 = const()[name = string("cache_77_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_77_cast_fp16 = slice_by_index(begin = cache_77_begin_0, end = cache_77_end_0, end_mask = cache_77_end_mask_0, squeeze_mask = cache_77_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_77_cast_fp16")];
+            tensor<int32, [4]> cache_79_begin_0 = const()[name = string("cache_79_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
+            tensor<int32, [4]> cache_79_end_0 = const()[name = string("cache_79_end_0"), val = tensor<int32, [4]>([20, 1, 1024, 8])];
+            tensor<bool, [4]> cache_79_end_mask_0 = const()[name = string("cache_79_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_79_squeeze_mask_0 = const()[name = string("cache_79_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_79_cast_fp16 = slice_by_index(begin = cache_79_begin_0, end = cache_79_end_0, end_mask = cache_79_end_mask_0, squeeze_mask = cache_79_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_79_cast_fp16")];
+            tensor<int32, [1]> input_1015_axes_0 = const()[name = string("input_1015_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383238336)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383240448)))];
+            tensor<fp16, [1, 28, 1024]> input_1015_cast_fp16 = layer_norm(axes = input_1015_axes_0, beta = encoder_layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_feed_forward1_weight_to_fp16, x = input_1013_cast_fp16)[name = string("input_1015_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383242560))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387436928))))[name = string("encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_19_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387445184)))];
+            tensor<fp16, [1, 28, 4096]> linear_172_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1015_cast_fp16)[name = string("linear_172_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_1019_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("input_1019_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(387453440))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391647808))))[name = string("encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391649920)))];
+            tensor<fp16, [1, 28, 1024]> linear_173_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1019_cast_fp16)[name = string("linear_173_cast_fp16")];
+            fp16 var_4502_to_fp16 = const()[name = string("op_4502_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_4503_cast_fp16 = mul(x = linear_173_cast_fp16, y = var_4502_to_fp16)[name = string("op_4503_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_1025_cast_fp16 = add(x = input_1013_cast_fp16, y = var_4503_cast_fp16)[name = string("input_1025_cast_fp16")];
+            tensor<int32, [1]> key_39_axes_0 = const()[name = string("key_39_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391652032)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391654144)))];
+            tensor<fp16, [1, 28, 1024]> key_39_cast_fp16 = layer_norm(axes = key_39_axes_0, beta = encoder_layers_19_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_self_att_weight_to_fp16, x = input_1025_cast_fp16)[name = string("key_39_cast_fp16")];
+            bool input_1027_interleave_0 = const()[name = string("input_1027_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_1027_cast_fp16 = concat(axis = var_68, interleave = input_1027_interleave_0, values = (cache_77_cast_fp16, key_39_cast_fp16))[name = string("input_1027_cast_fp16")];
+            tensor<int32, [3]> var_4525_begin_0 = const()[name = string("op_4525_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_4525_end_0 = const()[name = string("op_4525_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4525_end_mask_0 = const()[name = string("op_4525_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_4525_cast_fp16 = slice_by_index(begin = var_4525_begin_0, end = var_4525_end_0, end_mask = var_4525_end_mask_0, x = cache_77_cast_fp16)[name = string("op_4525_cast_fp16")];
+            bool var_4531_interleave_0 = const()[name = string("op_4531_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4531_cast_fp16 = concat(axis = var_68, interleave = var_4531_interleave_0, values = (var_4525_cast_fp16, key_39_cast_fp16))[name = string("op_4531_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391656256))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392704896))))[name = string("encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392707008)))];
+            tensor<fp16, [1, 28, 1024]> linear_174_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized, x = key_39_cast_fp16)[name = string("linear_174_cast_fp16")];
+            tensor<int32, [4]> var_4536 = const()[name = string("op_4536"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_115_cast_fp16 = reshape(shape = var_4536, x = linear_174_cast_fp16)[name = string("q_115_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392709120))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393757760))))[name = string("encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393759872)))];
+            tensor<fp16, [1, 70, 1024]> linear_175_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_175_cast_fp16")];
+            tensor<int32, [4]> var_4541 = const()[name = string("op_4541"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_77_cast_fp16 = reshape(shape = var_4541, x = linear_175_cast_fp16)[name = string("k_77_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393761984))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394810624))))[name = string("encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394812736)))];
+            tensor<fp16, [1, 70, 1024]> linear_176_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_176_cast_fp16")];
+            tensor<int32, [4]> var_4546 = const()[name = string("op_4546"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_39_cast_fp16 = reshape(shape = var_4546, x = linear_176_cast_fp16)[name = string("v_39_cast_fp16")];
+            tensor<int32, [4]> value_47_perm_0 = const()[name = string("value_47_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_19_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394814848)))];
+            tensor<fp16, [1, 28, 8, 128]> var_4559_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_u_to_fp16)[name = string("op_4559_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_19_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394816960)))];
+            tensor<fp16, [1, 28, 8, 128]> var_4561_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_v_to_fp16)[name = string("op_4561_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_39_perm_0 = const()[name = string("q_with_bias_v_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_501_transpose_x_0 = const()[name = string("x_501_transpose_x_0"), val = bool(false)];
+            bool x_501_transpose_y_0 = const()[name = string("x_501_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_4563_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394819072))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394961472))))[name = string("op_4563_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_39_cast_fp16 = transpose(perm = q_with_bias_v_39_perm_0, x = var_4561_cast_fp16)[name = string("transpose_191")];
+            tensor<fp16, [1, 8, 28, 139]> x_501_cast_fp16 = matmul(transpose_x = x_501_transpose_x_0, transpose_y = x_501_transpose_y_0, x = q_with_bias_v_39_cast_fp16, y = op_4563_to_fp16_quantized)[name = string("x_501_cast_fp16")];
+            tensor<int32, [8]> x_503_pad_0 = const()[name = string("x_503_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_503_mode_0 = const()[name = string("x_503_mode_0"), val = string("constant")];
+            fp16 const_326_to_fp16 = const()[name = string("const_326_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_503_cast_fp16 = pad(constant_val = const_326_to_fp16, mode = x_503_mode_0, pad = x_503_pad_0, x = x_501_cast_fp16)[name = string("x_503_cast_fp16")];
+            tensor<int32, [4]> var_4571 = const()[name = string("op_4571"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_505_cast_fp16 = reshape(shape = var_4571, x = x_503_cast_fp16)[name = string("x_505_cast_fp16")];
+            tensor<int32, [4]> var_4575_begin_0 = const()[name = string("op_4575_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4575_end_0 = const()[name = string("op_4575_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_4575_end_mask_0 = const()[name = string("op_4575_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_4575_cast_fp16 = slice_by_index(begin = var_4575_begin_0, end = var_4575_end_0, end_mask = var_4575_end_mask_0, x = x_505_cast_fp16)[name = string("op_4575_cast_fp16")];
+            tensor<int32, [4]> var_4576 = const()[name = string("op_4576"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_77_cast_fp16 = reshape(shape = var_4576, x = var_4575_cast_fp16)[name = string("matrix_bd_77_cast_fp16")];
+            bool matrix_ac_39_transpose_x_0 = const()[name = string("matrix_ac_39_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_39_transpose_y_0 = const()[name = string("matrix_ac_39_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_134_perm_0 = const()[name = string("transpose_134_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_135_perm_0 = const()[name = string("transpose_135_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_135 = transpose(perm = transpose_135_perm_0, x = k_77_cast_fp16)[name = string("transpose_189")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_134 = transpose(perm = transpose_134_perm_0, x = var_4559_cast_fp16)[name = string("transpose_190")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_39_cast_fp16 = matmul(transpose_x = matrix_ac_39_transpose_x_0, transpose_y = matrix_ac_39_transpose_y_0, x = transpose_134, y = transpose_135)[name = string("matrix_ac_39_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_79_begin_0 = const()[name = string("matrix_bd_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_79_end_0 = const()[name = string("matrix_bd_79_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_79_end_mask_0 = const()[name = string("matrix_bd_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_79_cast_fp16 = slice_by_index(begin = matrix_bd_79_begin_0, end = matrix_bd_79_end_0, end_mask = matrix_bd_79_end_mask_0, x = matrix_bd_77_cast_fp16)[name = string("matrix_bd_79_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_4585_cast_fp16 = add(x = matrix_ac_39_cast_fp16, y = matrix_bd_79_cast_fp16)[name = string("op_4585_cast_fp16")];
+            fp16 _inversed_scores_77_y_0_to_fp16 = const()[name = string("_inversed_scores_77_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_77_cast_fp16 = mul(x = var_4585_cast_fp16, y = _inversed_scores_77_y_0_to_fp16)[name = string("_inversed_scores_77_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_79_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_77_cast_fp16, cond = mask_11)[name = string("scores_79_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_4591_cast_fp16 = softmax(axis = var_59, x = scores_79_cast_fp16)[name = string("op_4591_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_1029_cast_fp16 = select(a = var_44_to_fp16, b = var_4591_cast_fp16, cond = mask_11)[name = string("input_1029_cast_fp16")];
+            bool x_507_transpose_x_0 = const()[name = string("x_507_transpose_x_0"), val = bool(false)];
+            bool x_507_transpose_y_0 = const()[name = string("x_507_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = v_39_cast_fp16)[name = string("transpose_188")];
+            tensor<fp16, [1, 8, 28, 128]> x_507_cast_fp16 = matmul(transpose_x = x_507_transpose_x_0, transpose_y = x_507_transpose_y_0, x = input_1029_cast_fp16, y = value_47_cast_fp16)[name = string("x_507_cast_fp16")];
+            tensor<int32, [4]> var_4595_perm_0 = const()[name = string("op_4595_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4596 = const()[name = string("op_4596"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_4595_cast_fp16 = transpose(perm = var_4595_perm_0, x = x_507_cast_fp16)[name = string("transpose_187")];
+            tensor<fp16, [1, 28, 1024]> input_1031_cast_fp16 = reshape(shape = var_4596, x = var_4595_cast_fp16)[name = string("input_1031_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394961856))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396010496))))[name = string("encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396012608)))];
+            tensor<fp16, [1, 28, 1024]> linear_178_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized, x = input_1031_cast_fp16)[name = string("linear_178_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_1035_cast_fp16 = add(x = input_1025_cast_fp16, y = linear_178_cast_fp16)[name = string("input_1035_cast_fp16")];
+            tensor<int32, [1]> x_511_axes_0 = const()[name = string("x_511_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396014720)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396016832)))];
+            tensor<fp16, [1, 28, 1024]> x_511_cast_fp16 = layer_norm(axes = x_511_axes_0, beta = encoder_layers_19_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_conv_weight_to_fp16, x = input_1035_cast_fp16)[name = string("x_511_cast_fp16")];
+            tensor<int32, [3]> input_1037_perm_0 = const()[name = string("input_1037_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1039_pad_type_0 = const()[name = string("input_1039_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1039_strides_0 = const()[name = string("input_1039_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1039_pad_0 = const()[name = string("input_1039_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1039_dilations_0 = const()[name = string("input_1039_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1039_groups_0 = const()[name = string("input_1039_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396018944))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398116160))))[name = string("encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_1037_cast_fp16 = transpose(perm = input_1037_perm_0, x = x_511_cast_fp16)[name = string("transpose_186")];
+            tensor<fp16, [1, 2048, 28]> input_1039_cast_fp16 = conv(dilations = input_1039_dilations_0, groups = input_1039_groups_0, pad = input_1039_pad_0, pad_type = input_1039_pad_type_0, strides = input_1039_strides_0, weight = encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1037_cast_fp16)[name = string("input_1039_cast_fp16")];
+            int32 x_513_split_num_splits_0 = const()[name = string("x_513_split_num_splits_0"), val = int32(2)];
+            int32 x_513_split_axis_0 = const()[name = string("x_513_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_513_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_513_split_cast_fp16_1 = split(axis = x_513_split_axis_0, num_splits = x_513_split_num_splits_0, x = input_1039_cast_fp16)[name = string("x_513_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_513_split_1_sigmoid_cast_fp16 = sigmoid(x = x_513_split_cast_fp16_1)[name = string("x_513_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_513_cast_fp16 = mul(x = x_513_split_cast_fp16_0, y = x_513_split_1_sigmoid_cast_fp16)[name = string("x_513_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_1041_cast_fp16 = select(a = var_44_to_fp16, b = x_513_cast_fp16, cond = var_575)[name = string("input_1041_cast_fp16")];
+            bool new_x_79_interleave_0 = const()[name = string("new_x_79_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_79_cast_fp16 = concat(axis = var_59, interleave = new_x_79_interleave_0, values = (cache_79_cast_fp16, input_1041_cast_fp16))[name = string("new_x_79_cast_fp16")];
+            tensor<int32, [3]> var_4635_begin_0 = const()[name = string("op_4635_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_4635_end_0 = const()[name = string("op_4635_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_4635_end_mask_0 = const()[name = string("op_4635_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4635_cast_fp16 = slice_by_index(begin = var_4635_begin_0, end = var_4635_end_0, end_mask = var_4635_end_mask_0, x = new_x_79_cast_fp16)[name = string("op_4635_cast_fp16")];
+            string x_515_pad_type_0 = const()[name = string("x_515_pad_type_0"), val = string("valid")];
+            int32 x_515_groups_0 = const()[name = string("x_515_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_515_strides_0 = const()[name = string("x_515_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_515_pad_0 = const()[name = string("x_515_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_515_dilations_0 = const()[name = string("x_515_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398120320))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398129600))))[name = string("encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_515_cast_fp16 = conv(dilations = x_515_dilations_0, groups = x_515_groups_0, pad = x_515_pad_0, pad_type = x_515_pad_type_0, strides = x_515_strides_0, weight = encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_79_cast_fp16)[name = string("x_515_cast_fp16")];
+            tensor<int32, [3]> input_1043_perm_0 = const()[name = string("input_1043_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_517_axes_0 = const()[name = string("x_517_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398131712)))];
+            tensor<fp16, [1024]> encoder_layers_19_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398133824)))];
+            tensor<fp16, [1, 28, 1024]> input_1043_cast_fp16 = transpose(perm = input_1043_perm_0, x = x_515_cast_fp16)[name = string("transpose_185")];
+            tensor<fp16, [1, 28, 1024]> x_517_cast_fp16 = layer_norm(axes = x_517_axes_0, beta = encoder_layers_19_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_conv_batch_norm_weight_to_fp16, x = input_1043_cast_fp16)[name = string("x_517_cast_fp16")];
+            tensor<int32, [3]> input_1045_perm_0 = const()[name = string("input_1045_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_1045_cast_fp16 = transpose(perm = input_1045_perm_0, x = x_517_cast_fp16)[name = string("transpose_184")];
+            tensor<fp16, [1, 1024, 28]> input_1047_cast_fp16 = silu(x = input_1045_cast_fp16)[name = string("input_1047_cast_fp16")];
+            string x_519_pad_type_0 = const()[name = string("x_519_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_519_strides_0 = const()[name = string("x_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_519_pad_0 = const()[name = string("x_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_519_dilations_0 = const()[name = string("x_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_519_groups_0 = const()[name = string("x_519_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398135936))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399184576))))[name = string("encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_519_cast_fp16 = conv(dilations = x_519_dilations_0, groups = x_519_groups_0, pad = x_519_pad_0, pad_type = x_519_pad_type_0, strides = x_519_strides_0, weight = encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1047_cast_fp16)[name = string("x_519_cast_fp16")];
+            tensor<int32, [3]> input_1049_perm_0 = const()[name = string("input_1049_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_1049_cast_fp16 = transpose(perm = input_1049_perm_0, x = x_519_cast_fp16)[name = string("transpose_183")];
+            tensor<fp16, [1, 28, 1024]> input_1051_cast_fp16 = add(x = input_1035_cast_fp16, y = input_1049_cast_fp16)[name = string("input_1051_cast_fp16")];
+            tensor<int32, [1]> input_1053_axes_0 = const()[name = string("input_1053_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399186688)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399188800)))];
+            tensor<fp16, [1, 28, 1024]> input_1053_cast_fp16 = layer_norm(axes = input_1053_axes_0, beta = encoder_layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_feed_forward2_weight_to_fp16, x = input_1051_cast_fp16)[name = string("input_1053_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399190912))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403385280))))[name = string("encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_19_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403393536)))];
+            tensor<fp16, [1, 28, 4096]> linear_179_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1053_cast_fp16)[name = string("linear_179_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_1057_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("input_1057_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403401792))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407596160))))[name = string("encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407598272)))];
+            tensor<fp16, [1, 28, 1024]> linear_180_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1057_cast_fp16)[name = string("linear_180_cast_fp16")];
+            fp16 var_4678_to_fp16 = const()[name = string("op_4678_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_4679_cast_fp16 = mul(x = linear_180_cast_fp16, y = var_4678_to_fp16)[name = string("op_4679_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_1063_cast_fp16 = add(x = input_1051_cast_fp16, y = var_4679_cast_fp16)[name = string("input_1063_cast_fp16")];
+            tensor<int32, [1]> input_1065_axes_0 = const()[name = string("input_1065_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407600384)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407602496)))];
+            tensor<fp16, [1, 28, 1024]> input_1065_cast_fp16 = layer_norm(axes = input_1065_axes_0, beta = encoder_layers_19_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_out_weight_to_fp16, x = input_1063_cast_fp16)[name = string("input_1065_cast_fp16")];
+            tensor<int32, [4]> cache_81_begin_0 = const()[name = string("cache_81_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
+            tensor<int32, [4]> cache_81_end_0 = const()[name = string("cache_81_end_0"), val = tensor<int32, [4]>([21, 1, 42, 1024])];
+            tensor<bool, [4]> cache_81_end_mask_0 = const()[name = string("cache_81_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_81_squeeze_mask_0 = const()[name = string("cache_81_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_81_cast_fp16 = slice_by_index(begin = cache_81_begin_0, end = cache_81_end_0, end_mask = cache_81_end_mask_0, squeeze_mask = cache_81_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_81_cast_fp16")];
+            tensor<int32, [4]> cache_83_begin_0 = const()[name = string("cache_83_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
+            tensor<int32, [4]> cache_83_end_0 = const()[name = string("cache_83_end_0"), val = tensor<int32, [4]>([21, 1, 1024, 8])];
+            tensor<bool, [4]> cache_83_end_mask_0 = const()[name = string("cache_83_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_83_squeeze_mask_0 = const()[name = string("cache_83_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_83_cast_fp16 = slice_by_index(begin = cache_83_begin_0, end = cache_83_end_0, end_mask = cache_83_end_mask_0, squeeze_mask = cache_83_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_83_cast_fp16")];
+            tensor<int32, [1]> input_1067_axes_0 = const()[name = string("input_1067_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407604608)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407606720)))];
+            tensor<fp16, [1, 28, 1024]> input_1067_cast_fp16 = layer_norm(axes = input_1067_axes_0, beta = encoder_layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_feed_forward1_weight_to_fp16, x = input_1065_cast_fp16)[name = string("input_1067_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407608832))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411803200))))[name = string("encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_20_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411811456)))];
+            tensor<fp16, [1, 28, 4096]> linear_181_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1067_cast_fp16)[name = string("linear_181_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_1071_cast_fp16 = silu(x = linear_181_cast_fp16)[name = string("input_1071_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411819712))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416014080))))[name = string("encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416016192)))];
+            tensor<fp16, [1, 28, 1024]> linear_182_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1071_cast_fp16)[name = string("linear_182_cast_fp16")];
+            fp16 var_4715_to_fp16 = const()[name = string("op_4715_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_4716_cast_fp16 = mul(x = linear_182_cast_fp16, y = var_4715_to_fp16)[name = string("op_4716_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_1077_cast_fp16 = add(x = input_1065_cast_fp16, y = var_4716_cast_fp16)[name = string("input_1077_cast_fp16")];
+            tensor<int32, [1]> key_41_axes_0 = const()[name = string("key_41_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416018304)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416020416)))];
+            tensor<fp16, [1, 28, 1024]> key_41_cast_fp16 = layer_norm(axes = key_41_axes_0, beta = encoder_layers_20_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_self_att_weight_to_fp16, x = input_1077_cast_fp16)[name = string("key_41_cast_fp16")];
+            bool input_1079_interleave_0 = const()[name = string("input_1079_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_1079_cast_fp16 = concat(axis = var_68, interleave = input_1079_interleave_0, values = (cache_81_cast_fp16, key_41_cast_fp16))[name = string("input_1079_cast_fp16")];
+            tensor<int32, [3]> var_4738_begin_0 = const()[name = string("op_4738_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_4738_end_0 = const()[name = string("op_4738_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4738_end_mask_0 = const()[name = string("op_4738_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_4738_cast_fp16 = slice_by_index(begin = var_4738_begin_0, end = var_4738_end_0, end_mask = var_4738_end_mask_0, x = cache_81_cast_fp16)[name = string("op_4738_cast_fp16")];
+            bool var_4744_interleave_0 = const()[name = string("op_4744_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4744_cast_fp16 = concat(axis = var_68, interleave = var_4744_interleave_0, values = (var_4738_cast_fp16, key_41_cast_fp16))[name = string("op_4744_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416022528))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417071168))))[name = string("encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417073280)))];
+            tensor<fp16, [1, 28, 1024]> linear_183_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized, x = key_41_cast_fp16)[name = string("linear_183_cast_fp16")];
+            tensor<int32, [4]> var_4749 = const()[name = string("op_4749"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_121_cast_fp16 = reshape(shape = var_4749, x = linear_183_cast_fp16)[name = string("q_121_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417075392))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418124032))))[name = string("encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418126144)))];
+            tensor<fp16, [1, 70, 1024]> linear_184_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_184_cast_fp16")];
+            tensor<int32, [4]> var_4754 = const()[name = string("op_4754"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_81_cast_fp16 = reshape(shape = var_4754, x = linear_184_cast_fp16)[name = string("k_81_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418128256))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419176896))))[name = string("encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419179008)))];
+            tensor<fp16, [1, 70, 1024]> linear_185_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_185_cast_fp16")];
+            tensor<int32, [4]> var_4759 = const()[name = string("op_4759"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_41_cast_fp16 = reshape(shape = var_4759, x = linear_185_cast_fp16)[name = string("v_41_cast_fp16")];
+            tensor<int32, [4]> value_49_perm_0 = const()[name = string("value_49_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_20_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419181120)))];
+            tensor<fp16, [1, 28, 8, 128]> var_4772_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_u_to_fp16)[name = string("op_4772_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_20_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419183232)))];
+            tensor<fp16, [1, 28, 8, 128]> var_4774_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_v_to_fp16)[name = string("op_4774_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_41_perm_0 = const()[name = string("q_with_bias_v_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_527_transpose_x_0 = const()[name = string("x_527_transpose_x_0"), val = bool(false)];
+            bool x_527_transpose_y_0 = const()[name = string("x_527_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_4776_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419185344))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419327744))))[name = string("op_4776_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_41_cast_fp16 = transpose(perm = q_with_bias_v_41_perm_0, x = var_4774_cast_fp16)[name = string("transpose_182")];
+            tensor<fp16, [1, 8, 28, 139]> x_527_cast_fp16 = matmul(transpose_x = x_527_transpose_x_0, transpose_y = x_527_transpose_y_0, x = q_with_bias_v_41_cast_fp16, y = op_4776_to_fp16_quantized)[name = string("x_527_cast_fp16")];
+            tensor<int32, [8]> x_529_pad_0 = const()[name = string("x_529_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_529_mode_0 = const()[name = string("x_529_mode_0"), val = string("constant")];
+            fp16 const_339_to_fp16 = const()[name = string("const_339_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_529_cast_fp16 = pad(constant_val = const_339_to_fp16, mode = x_529_mode_0, pad = x_529_pad_0, x = x_527_cast_fp16)[name = string("x_529_cast_fp16")];
+            tensor<int32, [4]> var_4784 = const()[name = string("op_4784"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_531_cast_fp16 = reshape(shape = var_4784, x = x_529_cast_fp16)[name = string("x_531_cast_fp16")];
+            tensor<int32, [4]> var_4788_begin_0 = const()[name = string("op_4788_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4788_end_0 = const()[name = string("op_4788_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_4788_end_mask_0 = const()[name = string("op_4788_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_4788_cast_fp16 = slice_by_index(begin = var_4788_begin_0, end = var_4788_end_0, end_mask = var_4788_end_mask_0, x = x_531_cast_fp16)[name = string("op_4788_cast_fp16")];
+            tensor<int32, [4]> var_4789 = const()[name = string("op_4789"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_81_cast_fp16 = reshape(shape = var_4789, x = var_4788_cast_fp16)[name = string("matrix_bd_81_cast_fp16")];
+            bool matrix_ac_41_transpose_x_0 = const()[name = string("matrix_ac_41_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_41_transpose_y_0 = const()[name = string("matrix_ac_41_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_136_perm_0 = const()[name = string("transpose_136_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_137_perm_0 = const()[name = string("transpose_137_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_137 = transpose(perm = transpose_137_perm_0, x = k_81_cast_fp16)[name = string("transpose_180")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_136 = transpose(perm = transpose_136_perm_0, x = var_4772_cast_fp16)[name = string("transpose_181")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_41_cast_fp16 = matmul(transpose_x = matrix_ac_41_transpose_x_0, transpose_y = matrix_ac_41_transpose_y_0, x = transpose_136, y = transpose_137)[name = string("matrix_ac_41_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_83_begin_0 = const()[name = string("matrix_bd_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_83_end_0 = const()[name = string("matrix_bd_83_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_83_end_mask_0 = const()[name = string("matrix_bd_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_83_cast_fp16 = slice_by_index(begin = matrix_bd_83_begin_0, end = matrix_bd_83_end_0, end_mask = matrix_bd_83_end_mask_0, x = matrix_bd_81_cast_fp16)[name = string("matrix_bd_83_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_4798_cast_fp16 = add(x = matrix_ac_41_cast_fp16, y = matrix_bd_83_cast_fp16)[name = string("op_4798_cast_fp16")];
+            fp16 _inversed_scores_81_y_0_to_fp16 = const()[name = string("_inversed_scores_81_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_81_cast_fp16 = mul(x = var_4798_cast_fp16, y = _inversed_scores_81_y_0_to_fp16)[name = string("_inversed_scores_81_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_83_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_81_cast_fp16, cond = mask_11)[name = string("scores_83_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_4804_cast_fp16 = softmax(axis = var_59, x = scores_83_cast_fp16)[name = string("op_4804_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_1081_cast_fp16 = select(a = var_44_to_fp16, b = var_4804_cast_fp16, cond = mask_11)[name = string("input_1081_cast_fp16")];
+            bool x_533_transpose_x_0 = const()[name = string("x_533_transpose_x_0"), val = bool(false)];
+            bool x_533_transpose_y_0 = const()[name = string("x_533_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = v_41_cast_fp16)[name = string("transpose_179")];
+            tensor<fp16, [1, 8, 28, 128]> x_533_cast_fp16 = matmul(transpose_x = x_533_transpose_x_0, transpose_y = x_533_transpose_y_0, x = input_1081_cast_fp16, y = value_49_cast_fp16)[name = string("x_533_cast_fp16")];
+            tensor<int32, [4]> var_4808_perm_0 = const()[name = string("op_4808_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4809 = const()[name = string("op_4809"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_4808_cast_fp16 = transpose(perm = var_4808_perm_0, x = x_533_cast_fp16)[name = string("transpose_178")];
+            tensor<fp16, [1, 28, 1024]> input_1083_cast_fp16 = reshape(shape = var_4809, x = var_4808_cast_fp16)[name = string("input_1083_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419328128))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420376768))))[name = string("encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420378880)))];
+            tensor<fp16, [1, 28, 1024]> linear_187_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized, x = input_1083_cast_fp16)[name = string("linear_187_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_1087_cast_fp16 = add(x = input_1077_cast_fp16, y = linear_187_cast_fp16)[name = string("input_1087_cast_fp16")];
+            tensor<int32, [1]> x_537_axes_0 = const()[name = string("x_537_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420380992)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420383104)))];
+            tensor<fp16, [1, 28, 1024]> x_537_cast_fp16 = layer_norm(axes = x_537_axes_0, beta = encoder_layers_20_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_conv_weight_to_fp16, x = input_1087_cast_fp16)[name = string("x_537_cast_fp16")];
+            tensor<int32, [3]> input_1089_perm_0 = const()[name = string("input_1089_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1091_pad_type_0 = const()[name = string("input_1091_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1091_strides_0 = const()[name = string("input_1091_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1091_pad_0 = const()[name = string("input_1091_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1091_dilations_0 = const()[name = string("input_1091_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1091_groups_0 = const()[name = string("input_1091_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420385216))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422482432))))[name = string("encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_1089_cast_fp16 = transpose(perm = input_1089_perm_0, x = x_537_cast_fp16)[name = string("transpose_177")];
+            tensor<fp16, [1, 2048, 28]> input_1091_cast_fp16 = conv(dilations = input_1091_dilations_0, groups = input_1091_groups_0, pad = input_1091_pad_0, pad_type = input_1091_pad_type_0, strides = input_1091_strides_0, weight = encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1089_cast_fp16)[name = string("input_1091_cast_fp16")];
+            int32 x_539_split_num_splits_0 = const()[name = string("x_539_split_num_splits_0"), val = int32(2)];
+            int32 x_539_split_axis_0 = const()[name = string("x_539_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_539_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_539_split_cast_fp16_1 = split(axis = x_539_split_axis_0, num_splits = x_539_split_num_splits_0, x = input_1091_cast_fp16)[name = string("x_539_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_539_split_1_sigmoid_cast_fp16 = sigmoid(x = x_539_split_cast_fp16_1)[name = string("x_539_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_539_cast_fp16 = mul(x = x_539_split_cast_fp16_0, y = x_539_split_1_sigmoid_cast_fp16)[name = string("x_539_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_1093_cast_fp16 = select(a = var_44_to_fp16, b = x_539_cast_fp16, cond = var_575)[name = string("input_1093_cast_fp16")];
+            bool new_x_83_interleave_0 = const()[name = string("new_x_83_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_83_cast_fp16 = concat(axis = var_59, interleave = new_x_83_interleave_0, values = (cache_83_cast_fp16, input_1093_cast_fp16))[name = string("new_x_83_cast_fp16")];
+            tensor<int32, [3]> var_4848_begin_0 = const()[name = string("op_4848_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_4848_end_0 = const()[name = string("op_4848_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_4848_end_mask_0 = const()[name = string("op_4848_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4848_cast_fp16 = slice_by_index(begin = var_4848_begin_0, end = var_4848_end_0, end_mask = var_4848_end_mask_0, x = new_x_83_cast_fp16)[name = string("op_4848_cast_fp16")];
+            string x_541_pad_type_0 = const()[name = string("x_541_pad_type_0"), val = string("valid")];
+            int32 x_541_groups_0 = const()[name = string("x_541_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_541_strides_0 = const()[name = string("x_541_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_541_pad_0 = const()[name = string("x_541_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_541_dilations_0 = const()[name = string("x_541_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422486592))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422495872))))[name = string("encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_541_cast_fp16 = conv(dilations = x_541_dilations_0, groups = x_541_groups_0, pad = x_541_pad_0, pad_type = x_541_pad_type_0, strides = x_541_strides_0, weight = encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_83_cast_fp16)[name = string("x_541_cast_fp16")];
+            tensor<int32, [3]> input_1095_perm_0 = const()[name = string("input_1095_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_543_axes_0 = const()[name = string("x_543_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422497984)))];
+            tensor<fp16, [1024]> encoder_layers_20_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422500096)))];
+            tensor<fp16, [1, 28, 1024]> input_1095_cast_fp16 = transpose(perm = input_1095_perm_0, x = x_541_cast_fp16)[name = string("transpose_176")];
+            tensor<fp16, [1, 28, 1024]> x_543_cast_fp16 = layer_norm(axes = x_543_axes_0, beta = encoder_layers_20_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_conv_batch_norm_weight_to_fp16, x = input_1095_cast_fp16)[name = string("x_543_cast_fp16")];
+            tensor<int32, [3]> input_1097_perm_0 = const()[name = string("input_1097_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_1097_cast_fp16 = transpose(perm = input_1097_perm_0, x = x_543_cast_fp16)[name = string("transpose_175")];
+            tensor<fp16, [1, 1024, 28]> input_1099_cast_fp16 = silu(x = input_1097_cast_fp16)[name = string("input_1099_cast_fp16")];
+            string x_545_pad_type_0 = const()[name = string("x_545_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_545_strides_0 = const()[name = string("x_545_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_545_pad_0 = const()[name = string("x_545_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_545_dilations_0 = const()[name = string("x_545_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_545_groups_0 = const()[name = string("x_545_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422502208))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423550848))))[name = string("encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_545_cast_fp16 = conv(dilations = x_545_dilations_0, groups = x_545_groups_0, pad = x_545_pad_0, pad_type = x_545_pad_type_0, strides = x_545_strides_0, weight = encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1099_cast_fp16)[name = string("x_545_cast_fp16")];
+            tensor<int32, [3]> input_1101_perm_0 = const()[name = string("input_1101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_1101_cast_fp16 = transpose(perm = input_1101_perm_0, x = x_545_cast_fp16)[name = string("transpose_174")];
+            tensor<fp16, [1, 28, 1024]> input_1103_cast_fp16 = add(x = input_1087_cast_fp16, y = input_1101_cast_fp16)[name = string("input_1103_cast_fp16")];
+            tensor<int32, [1]> input_1105_axes_0 = const()[name = string("input_1105_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423552960)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423555072)))];
+            tensor<fp16, [1, 28, 1024]> input_1105_cast_fp16 = layer_norm(axes = input_1105_axes_0, beta = encoder_layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_feed_forward2_weight_to_fp16, x = input_1103_cast_fp16)[name = string("input_1105_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423557184))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427751552))))[name = string("encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_20_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427759808)))];
+            tensor<fp16, [1, 28, 4096]> linear_188_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1105_cast_fp16)[name = string("linear_188_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_1109_cast_fp16 = silu(x = linear_188_cast_fp16)[name = string("input_1109_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427768064))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431962432))))[name = string("encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431964544)))];
+            tensor<fp16, [1, 28, 1024]> linear_189_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1109_cast_fp16)[name = string("linear_189_cast_fp16")];
+            fp16 var_4891_to_fp16 = const()[name = string("op_4891_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_4892_cast_fp16 = mul(x = linear_189_cast_fp16, y = var_4891_to_fp16)[name = string("op_4892_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_1115_cast_fp16 = add(x = input_1103_cast_fp16, y = var_4892_cast_fp16)[name = string("input_1115_cast_fp16")];
+            tensor<int32, [1]> input_1117_axes_0 = const()[name = string("input_1117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431966656)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431968768)))];
+            tensor<fp16, [1, 28, 1024]> input_1117_cast_fp16 = layer_norm(axes = input_1117_axes_0, beta = encoder_layers_20_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_out_weight_to_fp16, x = input_1115_cast_fp16)[name = string("input_1117_cast_fp16")];
+            tensor<int32, [4]> cache_85_begin_0 = const()[name = string("cache_85_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
+            tensor<int32, [4]> cache_85_end_0 = const()[name = string("cache_85_end_0"), val = tensor<int32, [4]>([22, 1, 42, 1024])];
+            tensor<bool, [4]> cache_85_end_mask_0 = const()[name = string("cache_85_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_85_squeeze_mask_0 = const()[name = string("cache_85_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_85_cast_fp16 = slice_by_index(begin = cache_85_begin_0, end = cache_85_end_0, end_mask = cache_85_end_mask_0, squeeze_mask = cache_85_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_85_cast_fp16")];
+            tensor<int32, [4]> cache_87_begin_0 = const()[name = string("cache_87_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
+            tensor<int32, [4]> cache_87_end_0 = const()[name = string("cache_87_end_0"), val = tensor<int32, [4]>([22, 1, 1024, 8])];
+            tensor<bool, [4]> cache_87_end_mask_0 = const()[name = string("cache_87_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_87_squeeze_mask_0 = const()[name = string("cache_87_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_87_cast_fp16 = slice_by_index(begin = cache_87_begin_0, end = cache_87_end_0, end_mask = cache_87_end_mask_0, squeeze_mask = cache_87_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_87_cast_fp16")];
+            tensor<int32, [1]> input_1119_axes_0 = const()[name = string("input_1119_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431970880)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431972992)))];
+            tensor<fp16, [1, 28, 1024]> input_1119_cast_fp16 = layer_norm(axes = input_1119_axes_0, beta = encoder_layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_feed_forward1_weight_to_fp16, x = input_1117_cast_fp16)[name = string("input_1119_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431975104))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436169472))))[name = string("encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_21_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436177728)))];
+            tensor<fp16, [1, 28, 4096]> linear_190_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1119_cast_fp16)[name = string("linear_190_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_1123_cast_fp16 = silu(x = linear_190_cast_fp16)[name = string("input_1123_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436185984))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440380352))))[name = string("encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440382464)))];
+            tensor<fp16, [1, 28, 1024]> linear_191_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1123_cast_fp16)[name = string("linear_191_cast_fp16")];
+            fp16 var_4928_to_fp16 = const()[name = string("op_4928_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_4929_cast_fp16 = mul(x = linear_191_cast_fp16, y = var_4928_to_fp16)[name = string("op_4929_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_1129_cast_fp16 = add(x = input_1117_cast_fp16, y = var_4929_cast_fp16)[name = string("input_1129_cast_fp16")];
+            tensor<int32, [1]> key_43_axes_0 = const()[name = string("key_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440384576)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440386688)))];
+            tensor<fp16, [1, 28, 1024]> key_43_cast_fp16 = layer_norm(axes = key_43_axes_0, beta = encoder_layers_21_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_self_att_weight_to_fp16, x = input_1129_cast_fp16)[name = string("key_43_cast_fp16")];
+            bool input_1131_interleave_0 = const()[name = string("input_1131_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_1131_cast_fp16 = concat(axis = var_68, interleave = input_1131_interleave_0, values = (cache_85_cast_fp16, key_43_cast_fp16))[name = string("input_1131_cast_fp16")];
+            tensor<int32, [3]> var_4951_begin_0 = const()[name = string("op_4951_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_4951_end_0 = const()[name = string("op_4951_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4951_end_mask_0 = const()[name = string("op_4951_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_4951_cast_fp16 = slice_by_index(begin = var_4951_begin_0, end = var_4951_end_0, end_mask = var_4951_end_mask_0, x = cache_85_cast_fp16)[name = string("op_4951_cast_fp16")];
+            bool var_4957_interleave_0 = const()[name = string("op_4957_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4957_cast_fp16 = concat(axis = var_68, interleave = var_4957_interleave_0, values = (var_4951_cast_fp16, key_43_cast_fp16))[name = string("op_4957_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440388800))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441437440))))[name = string("encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441439552)))];
+            tensor<fp16, [1, 28, 1024]> linear_192_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized, x = key_43_cast_fp16)[name = string("linear_192_cast_fp16")];
+            tensor<int32, [4]> var_4962 = const()[name = string("op_4962"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_127_cast_fp16 = reshape(shape = var_4962, x = linear_192_cast_fp16)[name = string("q_127_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441441664))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442490304))))[name = string("encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442492416)))];
+            tensor<fp16, [1, 70, 1024]> linear_193_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_193_cast_fp16")];
+            tensor<int32, [4]> var_4967 = const()[name = string("op_4967"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_85_cast_fp16 = reshape(shape = var_4967, x = linear_193_cast_fp16)[name = string("k_85_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442494528))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443543168))))[name = string("encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443545280)))];
+            tensor<fp16, [1, 70, 1024]> linear_194_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_194_cast_fp16")];
+            tensor<int32, [4]> var_4972 = const()[name = string("op_4972"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_43_cast_fp16 = reshape(shape = var_4972, x = linear_194_cast_fp16)[name = string("v_43_cast_fp16")];
+            tensor<int32, [4]> value_51_perm_0 = const()[name = string("value_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_21_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443547392)))];
+            tensor<fp16, [1, 28, 8, 128]> var_4985_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_u_to_fp16)[name = string("op_4985_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_21_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443549504)))];
+            tensor<fp16, [1, 28, 8, 128]> var_4987_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_v_to_fp16)[name = string("op_4987_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_43_perm_0 = const()[name = string("q_with_bias_v_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_553_transpose_x_0 = const()[name = string("x_553_transpose_x_0"), val = bool(false)];
+            bool x_553_transpose_y_0 = const()[name = string("x_553_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_4989_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443551616))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443694016))))[name = string("op_4989_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_43_cast_fp16 = transpose(perm = q_with_bias_v_43_perm_0, x = var_4987_cast_fp16)[name = string("transpose_173")];
+            tensor<fp16, [1, 8, 28, 139]> x_553_cast_fp16 = matmul(transpose_x = x_553_transpose_x_0, transpose_y = x_553_transpose_y_0, x = q_with_bias_v_43_cast_fp16, y = op_4989_to_fp16_quantized)[name = string("x_553_cast_fp16")];
+            tensor<int32, [8]> x_555_pad_0 = const()[name = string("x_555_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_555_mode_0 = const()[name = string("x_555_mode_0"), val = string("constant")];
+            fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_555_cast_fp16 = pad(constant_val = const_352_to_fp16, mode = x_555_mode_0, pad = x_555_pad_0, x = x_553_cast_fp16)[name = string("x_555_cast_fp16")];
+            tensor<int32, [4]> var_4997 = const()[name = string("op_4997"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_557_cast_fp16 = reshape(shape = var_4997, x = x_555_cast_fp16)[name = string("x_557_cast_fp16")];
+            tensor<int32, [4]> var_5001_begin_0 = const()[name = string("op_5001_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5001_end_0 = const()[name = string("op_5001_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_5001_end_mask_0 = const()[name = string("op_5001_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_5001_cast_fp16 = slice_by_index(begin = var_5001_begin_0, end = var_5001_end_0, end_mask = var_5001_end_mask_0, x = x_557_cast_fp16)[name = string("op_5001_cast_fp16")];
+            tensor<int32, [4]> var_5002 = const()[name = string("op_5002"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_85_cast_fp16 = reshape(shape = var_5002, x = var_5001_cast_fp16)[name = string("matrix_bd_85_cast_fp16")];
+            bool matrix_ac_43_transpose_x_0 = const()[name = string("matrix_ac_43_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_43_transpose_y_0 = const()[name = string("matrix_ac_43_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_138_perm_0 = const()[name = string("transpose_138_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_139_perm_0 = const()[name = string("transpose_139_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_139 = transpose(perm = transpose_139_perm_0, x = k_85_cast_fp16)[name = string("transpose_171")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_138 = transpose(perm = transpose_138_perm_0, x = var_4985_cast_fp16)[name = string("transpose_172")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_43_cast_fp16 = matmul(transpose_x = matrix_ac_43_transpose_x_0, transpose_y = matrix_ac_43_transpose_y_0, x = transpose_138, y = transpose_139)[name = string("matrix_ac_43_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_87_begin_0 = const()[name = string("matrix_bd_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_87_end_0 = const()[name = string("matrix_bd_87_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_87_end_mask_0 = const()[name = string("matrix_bd_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_87_cast_fp16 = slice_by_index(begin = matrix_bd_87_begin_0, end = matrix_bd_87_end_0, end_mask = matrix_bd_87_end_mask_0, x = matrix_bd_85_cast_fp16)[name = string("matrix_bd_87_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_5011_cast_fp16 = add(x = matrix_ac_43_cast_fp16, y = matrix_bd_87_cast_fp16)[name = string("op_5011_cast_fp16")];
+            fp16 _inversed_scores_85_y_0_to_fp16 = const()[name = string("_inversed_scores_85_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_85_cast_fp16 = mul(x = var_5011_cast_fp16, y = _inversed_scores_85_y_0_to_fp16)[name = string("_inversed_scores_85_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_87_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_85_cast_fp16, cond = mask_11)[name = string("scores_87_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_5017_cast_fp16 = softmax(axis = var_59, x = scores_87_cast_fp16)[name = string("op_5017_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_1133_cast_fp16 = select(a = var_44_to_fp16, b = var_5017_cast_fp16, cond = mask_11)[name = string("input_1133_cast_fp16")];
+            bool x_559_transpose_x_0 = const()[name = string("x_559_transpose_x_0"), val = bool(false)];
+            bool x_559_transpose_y_0 = const()[name = string("x_559_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = v_43_cast_fp16)[name = string("transpose_170")];
+            tensor<fp16, [1, 8, 28, 128]> x_559_cast_fp16 = matmul(transpose_x = x_559_transpose_x_0, transpose_y = x_559_transpose_y_0, x = input_1133_cast_fp16, y = value_51_cast_fp16)[name = string("x_559_cast_fp16")];
+            tensor<int32, [4]> var_5021_perm_0 = const()[name = string("op_5021_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5022 = const()[name = string("op_5022"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_5021_cast_fp16 = transpose(perm = var_5021_perm_0, x = x_559_cast_fp16)[name = string("transpose_169")];
+            tensor<fp16, [1, 28, 1024]> input_1135_cast_fp16 = reshape(shape = var_5022, x = var_5021_cast_fp16)[name = string("input_1135_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443694400)))];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445791616)))];
+            tensor<fp16, [1, 28, 1024]> linear_196_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_out_weight_to_fp16, x = input_1135_cast_fp16)[name = string("linear_196_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_1139_cast_fp16 = add(x = input_1129_cast_fp16, y = linear_196_cast_fp16)[name = string("input_1139_cast_fp16")];
+            tensor<int32, [1]> x_563_axes_0 = const()[name = string("x_563_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445793728)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445795840)))];
+            tensor<fp16, [1, 28, 1024]> x_563_cast_fp16 = layer_norm(axes = x_563_axes_0, beta = encoder_layers_21_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_conv_weight_to_fp16, x = input_1139_cast_fp16)[name = string("x_563_cast_fp16")];
+            tensor<int32, [3]> input_1141_perm_0 = const()[name = string("input_1141_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1143_pad_type_0 = const()[name = string("input_1143_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1143_strides_0 = const()[name = string("input_1143_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1143_pad_0 = const()[name = string("input_1143_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1143_dilations_0 = const()[name = string("input_1143_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1143_groups_0 = const()[name = string("input_1143_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445797952))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447895168))))[name = string("encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_1141_cast_fp16 = transpose(perm = input_1141_perm_0, x = x_563_cast_fp16)[name = string("transpose_168")];
+            tensor<fp16, [1, 2048, 28]> input_1143_cast_fp16 = conv(dilations = input_1143_dilations_0, groups = input_1143_groups_0, pad = input_1143_pad_0, pad_type = input_1143_pad_type_0, strides = input_1143_strides_0, weight = encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1141_cast_fp16)[name = string("input_1143_cast_fp16")];
+            int32 x_565_split_num_splits_0 = const()[name = string("x_565_split_num_splits_0"), val = int32(2)];
+            int32 x_565_split_axis_0 = const()[name = string("x_565_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_565_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_565_split_cast_fp16_1 = split(axis = x_565_split_axis_0, num_splits = x_565_split_num_splits_0, x = input_1143_cast_fp16)[name = string("x_565_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_565_split_1_sigmoid_cast_fp16 = sigmoid(x = x_565_split_cast_fp16_1)[name = string("x_565_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_565_cast_fp16 = mul(x = x_565_split_cast_fp16_0, y = x_565_split_1_sigmoid_cast_fp16)[name = string("x_565_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_1145_cast_fp16 = select(a = var_44_to_fp16, b = x_565_cast_fp16, cond = var_575)[name = string("input_1145_cast_fp16")];
+            bool new_x_87_interleave_0 = const()[name = string("new_x_87_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_87_cast_fp16 = concat(axis = var_59, interleave = new_x_87_interleave_0, values = (cache_87_cast_fp16, input_1145_cast_fp16))[name = string("new_x_87_cast_fp16")];
+            tensor<int32, [3]> var_5061_begin_0 = const()[name = string("op_5061_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_5061_end_0 = const()[name = string("op_5061_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_5061_end_mask_0 = const()[name = string("op_5061_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_5061_cast_fp16 = slice_by_index(begin = var_5061_begin_0, end = var_5061_end_0, end_mask = var_5061_end_mask_0, x = new_x_87_cast_fp16)[name = string("op_5061_cast_fp16")];
+            string x_567_pad_type_0 = const()[name = string("x_567_pad_type_0"), val = string("valid")];
+            int32 x_567_groups_0 = const()[name = string("x_567_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_567_strides_0 = const()[name = string("x_567_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_567_pad_0 = const()[name = string("x_567_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_567_dilations_0 = const()[name = string("x_567_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447899328))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447908608))))[name = string("encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_567_cast_fp16 = conv(dilations = x_567_dilations_0, groups = x_567_groups_0, pad = x_567_pad_0, pad_type = x_567_pad_type_0, strides = x_567_strides_0, weight = encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_87_cast_fp16)[name = string("x_567_cast_fp16")];
+            tensor<int32, [3]> input_1147_perm_0 = const()[name = string("input_1147_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_569_axes_0 = const()[name = string("x_569_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447910720)))];
+            tensor<fp16, [1024]> encoder_layers_21_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447912832)))];
+            tensor<fp16, [1, 28, 1024]> input_1147_cast_fp16 = transpose(perm = input_1147_perm_0, x = x_567_cast_fp16)[name = string("transpose_167")];
+            tensor<fp16, [1, 28, 1024]> x_569_cast_fp16 = layer_norm(axes = x_569_axes_0, beta = encoder_layers_21_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_conv_batch_norm_weight_to_fp16, x = input_1147_cast_fp16)[name = string("x_569_cast_fp16")];
+            tensor<int32, [3]> input_1149_perm_0 = const()[name = string("input_1149_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_1149_cast_fp16 = transpose(perm = input_1149_perm_0, x = x_569_cast_fp16)[name = string("transpose_166")];
+            tensor<fp16, [1, 1024, 28]> input_1151_cast_fp16 = silu(x = input_1149_cast_fp16)[name = string("input_1151_cast_fp16")];
+            string x_571_pad_type_0 = const()[name = string("x_571_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_571_strides_0 = const()[name = string("x_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_571_pad_0 = const()[name = string("x_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_571_dilations_0 = const()[name = string("x_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_571_groups_0 = const()[name = string("x_571_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447914944))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448963584))))[name = string("encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_571_cast_fp16 = conv(dilations = x_571_dilations_0, groups = x_571_groups_0, pad = x_571_pad_0, pad_type = x_571_pad_type_0, strides = x_571_strides_0, weight = encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1151_cast_fp16)[name = string("x_571_cast_fp16")];
+            tensor<int32, [3]> input_1153_perm_0 = const()[name = string("input_1153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_1153_cast_fp16 = transpose(perm = input_1153_perm_0, x = x_571_cast_fp16)[name = string("transpose_165")];
+            tensor<fp16, [1, 28, 1024]> input_1155_cast_fp16 = add(x = input_1139_cast_fp16, y = input_1153_cast_fp16)[name = string("input_1155_cast_fp16")];
+            tensor<int32, [1]> input_1157_axes_0 = const()[name = string("input_1157_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448965696)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448967808)))];
+            tensor<fp16, [1, 28, 1024]> input_1157_cast_fp16 = layer_norm(axes = input_1157_axes_0, beta = encoder_layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_feed_forward2_weight_to_fp16, x = input_1155_cast_fp16)[name = string("input_1157_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_21_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448969920)))];
+            tensor<fp16, [4096]> encoder_layers_21_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457358592)))];
+            tensor<fp16, [1, 28, 4096]> linear_197_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear1_weight_to_fp16, x = input_1157_cast_fp16)[name = string("linear_197_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_1161_cast_fp16 = silu(x = linear_197_cast_fp16)[name = string("input_1161_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_21_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457366848)))];
+            tensor<fp16, [1024]> encoder_layers_21_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465755520)))];
+            tensor<fp16, [1, 28, 1024]> linear_198_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear2_weight_to_fp16, x = input_1161_cast_fp16)[name = string("linear_198_cast_fp16")];
+            fp16 var_5104_to_fp16 = const()[name = string("op_5104_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_5105_cast_fp16 = mul(x = linear_198_cast_fp16, y = var_5104_to_fp16)[name = string("op_5105_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_1167_cast_fp16 = add(x = input_1155_cast_fp16, y = var_5105_cast_fp16)[name = string("input_1167_cast_fp16")];
+            tensor<int32, [1]> input_1169_axes_0 = const()[name = string("input_1169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465757632)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465759744)))];
+            tensor<fp16, [1, 28, 1024]> input_1169_cast_fp16 = layer_norm(axes = input_1169_axes_0, beta = encoder_layers_21_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_out_weight_to_fp16, x = input_1167_cast_fp16)[name = string("input_1169_cast_fp16")];
+            tensor<int32, [4]> cache_89_begin_0 = const()[name = string("cache_89_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
+            tensor<int32, [4]> cache_89_end_0 = const()[name = string("cache_89_end_0"), val = tensor<int32, [4]>([23, 1, 42, 1024])];
+            tensor<bool, [4]> cache_89_end_mask_0 = const()[name = string("cache_89_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_89_squeeze_mask_0 = const()[name = string("cache_89_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_89_cast_fp16 = slice_by_index(begin = cache_89_begin_0, end = cache_89_end_0, end_mask = cache_89_end_mask_0, squeeze_mask = cache_89_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_89_cast_fp16")];
+            tensor<int32, [4]> cache_91_begin_0 = const()[name = string("cache_91_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
+            tensor<int32, [4]> cache_91_end_0 = const()[name = string("cache_91_end_0"), val = tensor<int32, [4]>([23, 1, 1024, 8])];
+            tensor<bool, [4]> cache_91_end_mask_0 = const()[name = string("cache_91_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_91_squeeze_mask_0 = const()[name = string("cache_91_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_91_cast_fp16 = slice_by_index(begin = cache_91_begin_0, end = cache_91_end_0, end_mask = cache_91_end_mask_0, squeeze_mask = cache_91_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_91_cast_fp16")];
+            tensor<int32, [1]> input_1171_axes_0 = const()[name = string("input_1171_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465761856)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465763968)))];
+            tensor<fp16, [1, 28, 1024]> input_1171_cast_fp16 = layer_norm(axes = input_1171_axes_0, beta = encoder_layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_feed_forward1_weight_to_fp16, x = input_1169_cast_fp16)[name = string("input_1171_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_22_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465766080)))];
+            tensor<fp16, [4096]> encoder_layers_22_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474154752)))];
+            tensor<fp16, [1, 28, 4096]> linear_199_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear1_weight_to_fp16, x = input_1171_cast_fp16)[name = string("linear_199_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_1175_cast_fp16 = silu(x = linear_199_cast_fp16)[name = string("input_1175_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_22_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(474163008)))];
+            tensor<fp16, [1024]> encoder_layers_22_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482551680)))];
+            tensor<fp16, [1, 28, 1024]> linear_200_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear2_weight_to_fp16, x = input_1175_cast_fp16)[name = string("linear_200_cast_fp16")];
+            fp16 var_5141_to_fp16 = const()[name = string("op_5141_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_5142_cast_fp16 = mul(x = linear_200_cast_fp16, y = var_5141_to_fp16)[name = string("op_5142_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_1181_cast_fp16 = add(x = input_1169_cast_fp16, y = var_5142_cast_fp16)[name = string("input_1181_cast_fp16")];
+            tensor<int32, [1]> key_45_axes_0 = const()[name = string("key_45_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482553792)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482555904)))];
+            tensor<fp16, [1, 28, 1024]> key_45_cast_fp16 = layer_norm(axes = key_45_axes_0, beta = encoder_layers_22_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_self_att_weight_to_fp16, x = input_1181_cast_fp16)[name = string("key_45_cast_fp16")];
+            bool input_1183_interleave_0 = const()[name = string("input_1183_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_1183_cast_fp16 = concat(axis = var_68, interleave = input_1183_interleave_0, values = (cache_89_cast_fp16, key_45_cast_fp16))[name = string("input_1183_cast_fp16")];
+            tensor<int32, [3]> var_5164_begin_0 = const()[name = string("op_5164_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_5164_end_0 = const()[name = string("op_5164_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_5164_end_mask_0 = const()[name = string("op_5164_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_5164_cast_fp16 = slice_by_index(begin = var_5164_begin_0, end = var_5164_end_0, end_mask = var_5164_end_mask_0, x = cache_89_cast_fp16)[name = string("op_5164_cast_fp16")];
+            bool var_5170_interleave_0 = const()[name = string("op_5170_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_5170_cast_fp16 = concat(axis = var_68, interleave = var_5170_interleave_0, values = (var_5164_cast_fp16, key_45_cast_fp16))[name = string("op_5170_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482558016)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484655232)))];
+            tensor<fp16, [1, 28, 1024]> linear_201_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_q_weight_to_fp16, x = key_45_cast_fp16)[name = string("linear_201_cast_fp16")];
+            tensor<int32, [4]> var_5175 = const()[name = string("op_5175"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_133_cast_fp16 = reshape(shape = var_5175, x = linear_201_cast_fp16)[name = string("q_133_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484657344)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486754560)))];
+            tensor<fp16, [1, 70, 1024]> linear_202_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_k_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_202_cast_fp16")];
+            tensor<int32, [4]> var_5180 = const()[name = string("op_5180"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_89_cast_fp16 = reshape(shape = var_5180, x = linear_202_cast_fp16)[name = string("k_89_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486756672)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488853888)))];
+            tensor<fp16, [1, 70, 1024]> linear_203_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_v_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_203_cast_fp16")];
+            tensor<int32, [4]> var_5185 = const()[name = string("op_5185"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_45_cast_fp16 = reshape(shape = var_5185, x = linear_203_cast_fp16)[name = string("v_45_cast_fp16")];
+            tensor<int32, [4]> value_53_perm_0 = const()[name = string("value_53_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_22_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488856000)))];
+            tensor<fp16, [1, 28, 8, 128]> var_5198_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_u_to_fp16)[name = string("op_5198_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_22_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488858112)))];
+            tensor<fp16, [1, 28, 8, 128]> var_5200_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_v_to_fp16)[name = string("op_5200_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_45_perm_0 = const()[name = string("q_with_bias_v_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_579_transpose_x_0 = const()[name = string("x_579_transpose_x_0"), val = bool(false)];
+            bool x_579_transpose_y_0 = const()[name = string("x_579_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_5202_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488860224))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(489002624))))[name = string("op_5202_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_45_cast_fp16 = transpose(perm = q_with_bias_v_45_perm_0, x = var_5200_cast_fp16)[name = string("transpose_164")];
+            tensor<fp16, [1, 8, 28, 139]> x_579_cast_fp16 = matmul(transpose_x = x_579_transpose_x_0, transpose_y = x_579_transpose_y_0, x = q_with_bias_v_45_cast_fp16, y = op_5202_to_fp16_quantized)[name = string("x_579_cast_fp16")];
+            tensor<int32, [8]> x_581_pad_0 = const()[name = string("x_581_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_581_mode_0 = const()[name = string("x_581_mode_0"), val = string("constant")];
+            fp16 const_365_to_fp16 = const()[name = string("const_365_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_581_cast_fp16 = pad(constant_val = const_365_to_fp16, mode = x_581_mode_0, pad = x_581_pad_0, x = x_579_cast_fp16)[name = string("x_581_cast_fp16")];
+            tensor<int32, [4]> var_5210 = const()[name = string("op_5210"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_583_cast_fp16 = reshape(shape = var_5210, x = x_581_cast_fp16)[name = string("x_583_cast_fp16")];
+            tensor<int32, [4]> var_5214_begin_0 = const()[name = string("op_5214_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5214_end_0 = const()[name = string("op_5214_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_5214_end_mask_0 = const()[name = string("op_5214_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_5214_cast_fp16 = slice_by_index(begin = var_5214_begin_0, end = var_5214_end_0, end_mask = var_5214_end_mask_0, x = x_583_cast_fp16)[name = string("op_5214_cast_fp16")];
+            tensor<int32, [4]> var_5215 = const()[name = string("op_5215"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_89_cast_fp16 = reshape(shape = var_5215, x = var_5214_cast_fp16)[name = string("matrix_bd_89_cast_fp16")];
+            bool matrix_ac_45_transpose_x_0 = const()[name = string("matrix_ac_45_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_45_transpose_y_0 = const()[name = string("matrix_ac_45_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_140_perm_0 = const()[name = string("transpose_140_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_141_perm_0 = const()[name = string("transpose_141_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_141 = transpose(perm = transpose_141_perm_0, x = k_89_cast_fp16)[name = string("transpose_162")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_140 = transpose(perm = transpose_140_perm_0, x = var_5198_cast_fp16)[name = string("transpose_163")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_45_cast_fp16 = matmul(transpose_x = matrix_ac_45_transpose_x_0, transpose_y = matrix_ac_45_transpose_y_0, x = transpose_140, y = transpose_141)[name = string("matrix_ac_45_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_91_begin_0 = const()[name = string("matrix_bd_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_91_end_0 = const()[name = string("matrix_bd_91_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_91_end_mask_0 = const()[name = string("matrix_bd_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_91_cast_fp16 = slice_by_index(begin = matrix_bd_91_begin_0, end = matrix_bd_91_end_0, end_mask = matrix_bd_91_end_mask_0, x = matrix_bd_89_cast_fp16)[name = string("matrix_bd_91_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_5224_cast_fp16 = add(x = matrix_ac_45_cast_fp16, y = matrix_bd_91_cast_fp16)[name = string("op_5224_cast_fp16")];
+            fp16 _inversed_scores_89_y_0_to_fp16 = const()[name = string("_inversed_scores_89_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_89_cast_fp16 = mul(x = var_5224_cast_fp16, y = _inversed_scores_89_y_0_to_fp16)[name = string("_inversed_scores_89_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_91_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_89_cast_fp16, cond = mask_11)[name = string("scores_91_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_5230_cast_fp16 = softmax(axis = var_59, x = scores_91_cast_fp16)[name = string("op_5230_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_1185_cast_fp16 = select(a = var_44_to_fp16, b = var_5230_cast_fp16, cond = mask_11)[name = string("input_1185_cast_fp16")];
+            bool x_585_transpose_x_0 = const()[name = string("x_585_transpose_x_0"), val = bool(false)];
+            bool x_585_transpose_y_0 = const()[name = string("x_585_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = v_45_cast_fp16)[name = string("transpose_161")];
+            tensor<fp16, [1, 8, 28, 128]> x_585_cast_fp16 = matmul(transpose_x = x_585_transpose_x_0, transpose_y = x_585_transpose_y_0, x = input_1185_cast_fp16, y = value_53_cast_fp16)[name = string("x_585_cast_fp16")];
+            tensor<int32, [4]> var_5234_perm_0 = const()[name = string("op_5234_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5235 = const()[name = string("op_5235"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_5234_cast_fp16 = transpose(perm = var_5234_perm_0, x = x_585_cast_fp16)[name = string("transpose_160")];
+            tensor<fp16, [1, 28, 1024]> input_1187_cast_fp16 = reshape(shape = var_5235, x = var_5234_cast_fp16)[name = string("input_1187_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(489003008)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491100224)))];
+            tensor<fp16, [1, 28, 1024]> linear_205_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_out_weight_to_fp16, x = input_1187_cast_fp16)[name = string("linear_205_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_1191_cast_fp16 = add(x = input_1181_cast_fp16, y = linear_205_cast_fp16)[name = string("input_1191_cast_fp16")];
+            tensor<int32, [1]> x_589_axes_0 = const()[name = string("x_589_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491102336)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491104448)))];
+            tensor<fp16, [1, 28, 1024]> x_589_cast_fp16 = layer_norm(axes = x_589_axes_0, beta = encoder_layers_22_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_conv_weight_to_fp16, x = input_1191_cast_fp16)[name = string("x_589_cast_fp16")];
+            tensor<int32, [3]> input_1193_perm_0 = const()[name = string("input_1193_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1195_pad_type_0 = const()[name = string("input_1195_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1195_strides_0 = const()[name = string("input_1195_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1195_pad_0 = const()[name = string("input_1195_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1195_dilations_0 = const()[name = string("input_1195_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1195_groups_0 = const()[name = string("input_1195_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(491106560))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493203776))))[name = string("encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_1193_cast_fp16 = transpose(perm = input_1193_perm_0, x = x_589_cast_fp16)[name = string("transpose_159")];
+            tensor<fp16, [1, 2048, 28]> input_1195_cast_fp16 = conv(dilations = input_1195_dilations_0, groups = input_1195_groups_0, pad = input_1195_pad_0, pad_type = input_1195_pad_type_0, strides = input_1195_strides_0, weight = encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1193_cast_fp16)[name = string("input_1195_cast_fp16")];
+            int32 x_591_split_num_splits_0 = const()[name = string("x_591_split_num_splits_0"), val = int32(2)];
+            int32 x_591_split_axis_0 = const()[name = string("x_591_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_591_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_591_split_cast_fp16_1 = split(axis = x_591_split_axis_0, num_splits = x_591_split_num_splits_0, x = input_1195_cast_fp16)[name = string("x_591_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_591_split_1_sigmoid_cast_fp16 = sigmoid(x = x_591_split_cast_fp16_1)[name = string("x_591_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_591_cast_fp16 = mul(x = x_591_split_cast_fp16_0, y = x_591_split_1_sigmoid_cast_fp16)[name = string("x_591_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_1197_cast_fp16 = select(a = var_44_to_fp16, b = x_591_cast_fp16, cond = var_575)[name = string("input_1197_cast_fp16")];
+            bool new_x_91_interleave_0 = const()[name = string("new_x_91_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_91_cast_fp16 = concat(axis = var_59, interleave = new_x_91_interleave_0, values = (cache_91_cast_fp16, input_1197_cast_fp16))[name = string("new_x_91_cast_fp16")];
+            tensor<int32, [3]> var_5274_begin_0 = const()[name = string("op_5274_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> var_5274_end_0 = const()[name = string("op_5274_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> var_5274_end_mask_0 = const()[name = string("op_5274_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_5274_cast_fp16 = slice_by_index(begin = var_5274_begin_0, end = var_5274_end_0, end_mask = var_5274_end_mask_0, x = new_x_91_cast_fp16)[name = string("op_5274_cast_fp16")];
+            string x_593_pad_type_0 = const()[name = string("x_593_pad_type_0"), val = string("valid")];
+            int32 x_593_groups_0 = const()[name = string("x_593_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_593_strides_0 = const()[name = string("x_593_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_593_pad_0 = const()[name = string("x_593_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_593_dilations_0 = const()[name = string("x_593_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493207936))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493217216))))[name = string("encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_593_cast_fp16 = conv(dilations = x_593_dilations_0, groups = x_593_groups_0, pad = x_593_pad_0, pad_type = x_593_pad_type_0, strides = x_593_strides_0, weight = encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_91_cast_fp16)[name = string("x_593_cast_fp16")];
+            tensor<int32, [3]> input_1199_perm_0 = const()[name = string("input_1199_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_595_axes_0 = const()[name = string("x_595_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493219328)))];
+            tensor<fp16, [1024]> encoder_layers_22_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493221440)))];
+            tensor<fp16, [1, 28, 1024]> input_1199_cast_fp16 = transpose(perm = input_1199_perm_0, x = x_593_cast_fp16)[name = string("transpose_158")];
+            tensor<fp16, [1, 28, 1024]> x_595_cast_fp16 = layer_norm(axes = x_595_axes_0, beta = encoder_layers_22_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_conv_batch_norm_weight_to_fp16, x = input_1199_cast_fp16)[name = string("x_595_cast_fp16")];
+            tensor<int32, [3]> input_1201_perm_0 = const()[name = string("input_1201_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_1201_cast_fp16 = transpose(perm = input_1201_perm_0, x = x_595_cast_fp16)[name = string("transpose_157")];
+            tensor<fp16, [1, 1024, 28]> input_1203_cast_fp16 = silu(x = input_1201_cast_fp16)[name = string("input_1203_cast_fp16")];
+            string x_597_pad_type_0 = const()[name = string("x_597_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_597_strides_0 = const()[name = string("x_597_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_597_pad_0 = const()[name = string("x_597_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_597_dilations_0 = const()[name = string("x_597_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_597_groups_0 = const()[name = string("x_597_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493223552))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494272192))))[name = string("encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_597_cast_fp16 = conv(dilations = x_597_dilations_0, groups = x_597_groups_0, pad = x_597_pad_0, pad_type = x_597_pad_type_0, strides = x_597_strides_0, weight = encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1203_cast_fp16)[name = string("x_597_cast_fp16")];
+            tensor<int32, [3]> input_1205_perm_0 = const()[name = string("input_1205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_1205_cast_fp16 = transpose(perm = input_1205_perm_0, x = x_597_cast_fp16)[name = string("transpose_156")];
+            tensor<fp16, [1, 28, 1024]> input_1207_cast_fp16 = add(x = input_1191_cast_fp16, y = input_1205_cast_fp16)[name = string("input_1207_cast_fp16")];
+            tensor<int32, [1]> input_1209_axes_0 = const()[name = string("input_1209_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494274304)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494276416)))];
+            tensor<fp16, [1, 28, 1024]> input_1209_cast_fp16 = layer_norm(axes = input_1209_axes_0, beta = encoder_layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_feed_forward2_weight_to_fp16, x = input_1207_cast_fp16)[name = string("input_1209_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_22_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494278528)))];
+            tensor<fp16, [4096]> encoder_layers_22_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502667200)))];
+            tensor<fp16, [1, 28, 4096]> linear_206_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear1_weight_to_fp16, x = input_1209_cast_fp16)[name = string("linear_206_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_1213_cast_fp16 = silu(x = linear_206_cast_fp16)[name = string("input_1213_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_22_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502675456)))];
+            tensor<fp16, [1024]> encoder_layers_22_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511064128)))];
+            tensor<fp16, [1, 28, 1024]> linear_207_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear2_weight_to_fp16, x = input_1213_cast_fp16)[name = string("linear_207_cast_fp16")];
+            fp16 var_5317_to_fp16 = const()[name = string("op_5317_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_5318_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_5317_to_fp16)[name = string("op_5318_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_1219_cast_fp16 = add(x = input_1207_cast_fp16, y = var_5318_cast_fp16)[name = string("input_1219_cast_fp16")];
+            tensor<int32, [1]> input_1221_axes_0 = const()[name = string("input_1221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511066240)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511068352)))];
+            tensor<fp16, [1, 28, 1024]> input_1221_cast_fp16 = layer_norm(axes = input_1221_axes_0, beta = encoder_layers_22_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_out_weight_to_fp16, x = input_1219_cast_fp16)[name = string("input_1221_cast_fp16")];
+            tensor<int32, [4]> cache_93_begin_0 = const()[name = string("cache_93_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
+            tensor<int32, [4]> cache_93_end_0 = const()[name = string("cache_93_end_0"), val = tensor<int32, [4]>([24, 1, 42, 1024])];
+            tensor<bool, [4]> cache_93_end_mask_0 = const()[name = string("cache_93_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_93_squeeze_mask_0 = const()[name = string("cache_93_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_93_cast_fp16 = slice_by_index(begin = cache_93_begin_0, end = cache_93_end_0, end_mask = cache_93_end_mask_0, squeeze_mask = cache_93_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_93_cast_fp16")];
+            tensor<int32, [4]> cache_begin_0 = const()[name = string("cache_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
+            tensor<int32, [4]> cache_end_0 = const()[name = string("cache_end_0"), val = tensor<int32, [4]>([24, 1, 1024, 8])];
+            tensor<bool, [4]> cache_end_mask_0 = const()[name = string("cache_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_squeeze_mask_0 = const()[name = string("cache_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_cast_fp16 = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_cast_fp16")];
+            tensor<int32, [1]> input_1223_axes_0 = const()[name = string("input_1223_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511070464)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511072576)))];
+            tensor<fp16, [1, 28, 1024]> input_1223_cast_fp16 = layer_norm(axes = input_1223_axes_0, beta = encoder_layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_feed_forward1_weight_to_fp16, x = input_1221_cast_fp16)[name = string("input_1223_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_23_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511074688)))];
+            tensor<fp16, [4096]> encoder_layers_23_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519463360)))];
+            tensor<fp16, [1, 28, 4096]> linear_208_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear1_weight_to_fp16, x = input_1223_cast_fp16)[name = string("linear_208_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_1227_cast_fp16 = silu(x = linear_208_cast_fp16)[name = string("input_1227_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_23_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519471616)))];
+            tensor<fp16, [1024]> encoder_layers_23_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527860288)))];
+            tensor<fp16, [1, 28, 1024]> linear_209_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear2_weight_to_fp16, x = input_1227_cast_fp16)[name = string("linear_209_cast_fp16")];
+            fp16 var_5354_to_fp16 = const()[name = string("op_5354_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_5355_cast_fp16 = mul(x = linear_209_cast_fp16, y = var_5354_to_fp16)[name = string("op_5355_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_1233_cast_fp16 = add(x = input_1221_cast_fp16, y = var_5355_cast_fp16)[name = string("input_1233_cast_fp16")];
+            tensor<int32, [1]> key_axes_0 = const()[name = string("key_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527862400)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527864512)))];
+            tensor<fp16, [1, 28, 1024]> key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_23_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_self_att_weight_to_fp16, x = input_1233_cast_fp16)[name = string("key_cast_fp16")];
+            bool input_1235_interleave_0 = const()[name = string("input_1235_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 70, 1024]> input_1235_cast_fp16 = concat(axis = var_68, interleave = input_1235_interleave_0, values = (cache_93_cast_fp16, key_cast_fp16))[name = string("input_1235_cast_fp16")];
+            tensor<int32, [3]> var_5377_begin_0 = const()[name = string("op_5377_begin_0"), val = tensor<int32, [3]>([0, 28, 0])];
+            tensor<int32, [3]> var_5377_end_0 = const()[name = string("op_5377_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_5377_end_mask_0 = const()[name = string("op_5377_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 14, 1024]> var_5377_cast_fp16 = slice_by_index(begin = var_5377_begin_0, end = var_5377_end_0, end_mask = var_5377_end_mask_0, x = cache_93_cast_fp16)[name = string("op_5377_cast_fp16")];
+            bool cache_last_channel_cur_interleave_0 = const()[name = string("cache_last_channel_cur_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> cache_last_channel_cur_cast_fp16 = concat(axis = var_68, interleave = cache_last_channel_cur_interleave_0, values = (var_5377_cast_fp16, key_cast_fp16))[name = string("cache_last_channel_cur_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527866624)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529963840)))];
+            tensor<fp16, [1, 28, 1024]> linear_210_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_q_weight_to_fp16, x = key_cast_fp16)[name = string("linear_210_cast_fp16")];
+            tensor<int32, [4]> var_5388 = const()[name = string("op_5388"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 28, 8, 128]> q_139_cast_fp16 = reshape(shape = var_5388, x = linear_210_cast_fp16)[name = string("q_139_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529965952)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(532063168)))];
+            tensor<fp16, [1, 70, 1024]> linear_211_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_k_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_211_cast_fp16")];
+            tensor<int32, [4]> var_5393 = const()[name = string("op_5393"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> k_93_cast_fp16 = reshape(shape = var_5393, x = linear_211_cast_fp16)[name = string("k_93_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(532065280)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534162496)))];
+            tensor<fp16, [1, 70, 1024]> linear_212_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_v_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_212_cast_fp16")];
+            tensor<int32, [4]> var_5398 = const()[name = string("op_5398"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 70, 8, 128]> v_cast_fp16 = reshape(shape = var_5398, x = linear_212_cast_fp16)[name = string("v_cast_fp16")];
+            tensor<int32, [4]> value_perm_0 = const()[name = string("value_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_23_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534164608)))];
+            tensor<fp16, [1, 28, 8, 128]> var_5411_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_u_to_fp16)[name = string("op_5411_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_23_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534166720)))];
+            tensor<fp16, [1, 28, 8, 128]> var_5413_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_v_to_fp16)[name = string("op_5413_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_perm_0 = const()[name = string("q_with_bias_v_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_605_transpose_x_0 = const()[name = string("x_605_transpose_x_0"), val = bool(false)];
+            bool x_605_transpose_y_0 = const()[name = string("x_605_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 139]> op_5415_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534168832))), scale = tensor<fp16, [1, 1, 1, 139]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534311232))))[name = string("op_5415_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 28, 128]> q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_5413_cast_fp16)[name = string("transpose_155")];
+            tensor<fp16, [1, 8, 28, 139]> x_605_cast_fp16 = matmul(transpose_x = x_605_transpose_x_0, transpose_y = x_605_transpose_y_0, x = q_with_bias_v_cast_fp16, y = op_5415_to_fp16_quantized)[name = string("x_605_cast_fp16")];
+            tensor<int32, [8]> x_607_pad_0 = const()[name = string("x_607_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_607_mode_0 = const()[name = string("x_607_mode_0"), val = string("constant")];
+            fp16 const_378_to_fp16 = const()[name = string("const_378_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 28, 140]> x_607_cast_fp16 = pad(constant_val = const_378_to_fp16, mode = x_607_mode_0, pad = x_607_pad_0, x = x_605_cast_fp16)[name = string("x_607_cast_fp16")];
+            tensor<int32, [4]> var_5423 = const()[name = string("op_5423"), val = tensor<int32, [4]>([1, 8, -1, 28])];
+            tensor<fp16, [1, 8, 140, 28]> x_609_cast_fp16 = reshape(shape = var_5423, x = x_607_cast_fp16)[name = string("x_609_cast_fp16")];
+            tensor<int32, [4]> var_5427_begin_0 = const()[name = string("op_5427_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5427_end_0 = const()[name = string("op_5427_end_0"), val = tensor<int32, [4]>([1, 8, 140, 28])];
+            tensor<bool, [4]> var_5427_end_mask_0 = const()[name = string("op_5427_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 139, 28]> var_5427_cast_fp16 = slice_by_index(begin = var_5427_begin_0, end = var_5427_end_0, end_mask = var_5427_end_mask_0, x = x_609_cast_fp16)[name = string("op_5427_cast_fp16")];
+            tensor<int32, [4]> var_5428 = const()[name = string("op_5428"), val = tensor<int32, [4]>([1, 8, 28, 139])];
+            tensor<fp16, [1, 8, 28, 139]> matrix_bd_93_cast_fp16 = reshape(shape = var_5428, x = var_5427_cast_fp16)[name = string("matrix_bd_93_cast_fp16")];
+            bool matrix_ac_transpose_x_0 = const()[name = string("matrix_ac_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_transpose_y_0 = const()[name = string("matrix_ac_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_142_perm_0 = const()[name = string("transpose_142_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_143_perm_0 = const()[name = string("transpose_143_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 70]> transpose_143 = transpose(perm = transpose_143_perm_0, x = k_93_cast_fp16)[name = string("transpose_153")];
+            tensor<fp16, [1, 8, 28, 128]> transpose_142 = transpose(perm = transpose_142_perm_0, x = var_5411_cast_fp16)[name = string("transpose_154")];
+            tensor<fp16, [1, 8, 28, 70]> matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_142, y = transpose_143)[name = string("matrix_ac_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_begin_0 = const()[name = string("matrix_bd_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_end_0 = const()[name = string("matrix_bd_end_0"), val = tensor<int32, [4]>([1, 8, 28, 70])];
+            tensor<bool, [4]> matrix_bd_end_mask_0 = const()[name = string("matrix_bd_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 28, 70]> matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_93_cast_fp16)[name = string("matrix_bd_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_5437_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = string("op_5437_cast_fp16")];
+            fp16 _inversed_scores_93_y_0_to_fp16 = const()[name = string("_inversed_scores_93_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 28, 70]> _inversed_scores_93_cast_fp16 = mul(x = var_5437_cast_fp16, y = _inversed_scores_93_y_0_to_fp16)[name = string("_inversed_scores_93_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> scores_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_93_cast_fp16, cond = mask_11)[name = string("scores_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> var_5443_cast_fp16 = softmax(axis = var_59, x = scores_cast_fp16)[name = string("op_5443_cast_fp16")];
+            tensor<fp16, [1, 8, 28, 70]> input_1237_cast_fp16 = select(a = var_44_to_fp16, b = var_5443_cast_fp16, cond = mask_11)[name = string("input_1237_cast_fp16")];
+            bool x_611_transpose_x_0 = const()[name = string("x_611_transpose_x_0"), val = bool(false)];
+            bool x_611_transpose_y_0 = const()[name = string("x_611_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 70, 128]> value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = string("transpose_152")];
+            tensor<fp16, [1, 8, 28, 128]> x_611_cast_fp16 = matmul(transpose_x = x_611_transpose_x_0, transpose_y = x_611_transpose_y_0, x = input_1237_cast_fp16, y = value_cast_fp16)[name = string("x_611_cast_fp16")];
+            tensor<int32, [4]> var_5447_perm_0 = const()[name = string("op_5447_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5448 = const()[name = string("op_5448"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 28, 8, 128]> var_5447_cast_fp16 = transpose(perm = var_5447_perm_0, x = x_611_cast_fp16)[name = string("transpose_151")];
+            tensor<fp16, [1, 28, 1024]> input_1239_cast_fp16 = reshape(shape = var_5448, x = var_5447_cast_fp16)[name = string("input_1239_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534311616)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536408832)))];
+            tensor<fp16, [1, 28, 1024]> linear_214_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_out_weight_to_fp16, x = input_1239_cast_fp16)[name = string("linear_214_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_1243_cast_fp16 = add(x = input_1233_cast_fp16, y = linear_214_cast_fp16)[name = string("input_1243_cast_fp16")];
+            tensor<int32, [1]> x_615_axes_0 = const()[name = string("x_615_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536410944)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536413056)))];
+            tensor<fp16, [1, 28, 1024]> x_615_cast_fp16 = layer_norm(axes = x_615_axes_0, beta = encoder_layers_23_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_conv_weight_to_fp16, x = input_1243_cast_fp16)[name = string("x_615_cast_fp16")];
+            tensor<int32, [3]> input_1245_perm_0 = const()[name = string("input_1245_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1247_pad_type_0 = const()[name = string("input_1247_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1247_strides_0 = const()[name = string("input_1247_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1247_pad_0 = const()[name = string("input_1247_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1247_dilations_0 = const()[name = string("input_1247_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1247_groups_0 = const()[name = string("input_1247_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536415168))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538512384))))[name = string("encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> input_1245_cast_fp16 = transpose(perm = input_1245_perm_0, x = x_615_cast_fp16)[name = string("transpose_150")];
+            tensor<fp16, [1, 2048, 28]> input_1247_cast_fp16 = conv(dilations = input_1247_dilations_0, groups = input_1247_groups_0, pad = input_1247_pad_0, pad_type = input_1247_pad_type_0, strides = input_1247_strides_0, weight = encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1245_cast_fp16)[name = string("input_1247_cast_fp16")];
+            int32 x_617_split_num_splits_0 = const()[name = string("x_617_split_num_splits_0"), val = int32(2)];
+            int32 x_617_split_axis_0 = const()[name = string("x_617_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 28]> x_617_split_cast_fp16_0, tensor<fp16, [1, 1024, 28]> x_617_split_cast_fp16_1 = split(axis = x_617_split_axis_0, num_splits = x_617_split_num_splits_0, x = input_1247_cast_fp16)[name = string("x_617_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_617_split_1_sigmoid_cast_fp16 = sigmoid(x = x_617_split_cast_fp16_1)[name = string("x_617_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> x_617_cast_fp16 = mul(x = x_617_split_cast_fp16_0, y = x_617_split_1_sigmoid_cast_fp16)[name = string("x_617_cast_fp16")];
+            tensor<fp16, [1, 1024, 28]> input_1249_cast_fp16 = select(a = var_44_to_fp16, b = x_617_cast_fp16, cond = var_575)[name = string("input_1249_cast_fp16")];
+            bool new_x_interleave_0 = const()[name = string("new_x_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 36]> new_x_cast_fp16 = concat(axis = var_59, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_1249_cast_fp16))[name = string("new_x_cast_fp16")];
+            tensor<int32, [3]> cache_last_time_cur_begin_0 = const()[name = string("cache_last_time_cur_begin_0"), val = tensor<int32, [3]>([0, 0, 28])];
+            tensor<int32, [3]> cache_last_time_cur_end_0 = const()[name = string("cache_last_time_cur_end_0"), val = tensor<int32, [3]>([1, 1024, 36])];
+            tensor<bool, [3]> cache_last_time_cur_end_mask_0 = const()[name = string("cache_last_time_cur_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> cache_last_time_cur_cast_fp16 = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = new_x_cast_fp16)[name = string("cache_last_time_cur_cast_fp16")];
+            string x_619_pad_type_0 = const()[name = string("x_619_pad_type_0"), val = string("valid")];
+            int32 x_619_groups_0 = const()[name = string("x_619_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_619_strides_0 = const()[name = string("x_619_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_619_pad_0 = const()[name = string("x_619_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_619_dilations_0 = const()[name = string("x_619_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538516544))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538525824))))[name = string("encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_619_cast_fp16 = conv(dilations = x_619_dilations_0, groups = x_619_groups_0, pad = x_619_pad_0, pad_type = x_619_pad_type_0, strides = x_619_strides_0, weight = encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_cast_fp16)[name = string("x_619_cast_fp16")];
+            tensor<int32, [3]> input_1251_perm_0 = const()[name = string("input_1251_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_621_axes_0 = const()[name = string("x_621_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538527936)))];
+            tensor<fp16, [1024]> encoder_layers_23_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538530048)))];
+            tensor<fp16, [1, 28, 1024]> input_1251_cast_fp16 = transpose(perm = input_1251_perm_0, x = x_619_cast_fp16)[name = string("transpose_149")];
+            tensor<fp16, [1, 28, 1024]> x_621_cast_fp16 = layer_norm(axes = x_621_axes_0, beta = encoder_layers_23_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_conv_batch_norm_weight_to_fp16, x = input_1251_cast_fp16)[name = string("x_621_cast_fp16")];
+            tensor<int32, [3]> input_1253_perm_0 = const()[name = string("input_1253_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 28]> input_1253_cast_fp16 = transpose(perm = input_1253_perm_0, x = x_621_cast_fp16)[name = string("transpose_148")];
+            tensor<fp16, [1, 1024, 28]> input_1255_cast_fp16 = silu(x = input_1253_cast_fp16)[name = string("input_1255_cast_fp16")];
+            string x_623_pad_type_0 = const()[name = string("x_623_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_623_strides_0 = const()[name = string("x_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_623_pad_0 = const()[name = string("x_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_623_dilations_0 = const()[name = string("x_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_623_groups_0 = const()[name = string("x_623_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538532160))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539580800))))[name = string("encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 28]> x_623_cast_fp16 = conv(dilations = x_623_dilations_0, groups = x_623_groups_0, pad = x_623_pad_0, pad_type = x_623_pad_type_0, strides = x_623_strides_0, weight = encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1255_cast_fp16)[name = string("x_623_cast_fp16")];
+            tensor<int32, [3]> input_1257_perm_0 = const()[name = string("input_1257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 28, 1024]> input_1257_cast_fp16 = transpose(perm = input_1257_perm_0, x = x_623_cast_fp16)[name = string("transpose_147")];
+            tensor<fp16, [1, 28, 1024]> input_1259_cast_fp16 = add(x = input_1243_cast_fp16, y = input_1257_cast_fp16)[name = string("input_1259_cast_fp16")];
+            tensor<int32, [1]> input_1261_axes_0 = const()[name = string("input_1261_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539582912)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539585024)))];
+            tensor<fp16, [1, 28, 1024]> input_1261_cast_fp16 = layer_norm(axes = input_1261_axes_0, beta = encoder_layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_feed_forward2_weight_to_fp16, x = input_1259_cast_fp16)[name = string("input_1261_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_23_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539587136)))];
+            tensor<fp16, [4096]> encoder_layers_23_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547975808)))];
+            tensor<fp16, [1, 28, 4096]> linear_215_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear1_weight_to_fp16, x = input_1261_cast_fp16)[name = string("linear_215_cast_fp16")];
+            tensor<fp16, [1, 28, 4096]> input_1265_cast_fp16 = silu(x = linear_215_cast_fp16)[name = string("input_1265_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_23_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547984064)))];
+            tensor<fp16, [1024]> encoder_layers_23_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(556372736)))];
+            tensor<fp16, [1, 28, 1024]> linear_216_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear2_weight_to_fp16, x = input_1265_cast_fp16)[name = string("linear_216_cast_fp16")];
+            fp16 var_5530_to_fp16 = const()[name = string("op_5530_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 28, 1024]> var_5531_cast_fp16 = mul(x = linear_216_cast_fp16, y = var_5530_to_fp16)[name = string("op_5531_cast_fp16")];
+            tensor<fp16, [1, 28, 1024]> input_1271_cast_fp16 = add(x = input_1259_cast_fp16, y = var_5531_cast_fp16)[name = string("input_1271_cast_fp16")];
+            tensor<int32, [1]> audio_signal_axes_0 = const()[name = string("audio_signal_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(556374848)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(556376960)))];
+            tensor<fp16, [1, 28, 1024]> audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_23_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_out_weight_to_fp16, x = input_1271_cast_fp16)[name = string("audio_signal_cast_fp16")];
+            int32 obj_5_axis_0 = const()[name = string("obj_5_axis_0"), val = int32(0)];
+            tensor<fp16, [24, 1, 42, 1024]> obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_484_cast_fp16, var_697_cast_fp16, var_910_cast_fp16, var_1123_cast_fp16, var_1336_cast_fp16, var_1549_cast_fp16, var_1762_cast_fp16, var_1975_cast_fp16, var_2188_cast_fp16, var_2401_cast_fp16, var_2614_cast_fp16, var_2827_cast_fp16, var_3040_cast_fp16, var_3253_cast_fp16, var_3466_cast_fp16, var_3679_cast_fp16, var_3892_cast_fp16, var_4105_cast_fp16, var_4318_cast_fp16, var_4531_cast_fp16, var_4744_cast_fp16, var_4957_cast_fp16, var_5170_cast_fp16, cache_last_channel_cur_cast_fp16))[name = string("obj_5_cast_fp16")];
+            int32 obj_7_axis_0 = const()[name = string("obj_7_axis_0"), val = int32(0)];
+            tensor<fp16, [24, 1, 1024, 8]> obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_588_cast_fp16, var_801_cast_fp16, var_1014_cast_fp16, var_1227_cast_fp16, var_1440_cast_fp16, var_1653_cast_fp16, var_1866_cast_fp16, var_2079_cast_fp16, var_2292_cast_fp16, var_2505_cast_fp16, var_2718_cast_fp16, var_2931_cast_fp16, var_3144_cast_fp16, var_3357_cast_fp16, var_3570_cast_fp16, var_3783_cast_fp16, var_3996_cast_fp16, var_4209_cast_fp16, var_4422_cast_fp16, var_4635_cast_fp16, var_4848_cast_fp16, var_5061_cast_fp16, var_5274_cast_fp16, cache_last_time_cur_cast_fp16))[name = string("obj_7_cast_fp16")];
+            tensor<int32, [1]> var_5547 = add(x = cache_len, y = max_audio_length_1)[name = string("op_5547")];
+            string var_5547_promoted_to_fp16_dtype_0 = const()[name = string("op_5547_promoted_to_fp16_dtype_0"), val = string("fp16")];
+            fp16 const_384_to_fp16 = const()[name = string("const_384_to_fp16"), val = fp16(-inf)];
+            fp16 var_49_promoted_to_fp16 = const()[name = string("op_49_promoted_to_fp16"), val = fp16(0x1.5p+5)];
+            tensor<fp16, [1]> var_5547_to_fp16 = cast(dtype = var_5547_promoted_to_fp16_dtype_0, x = var_5547)[name = string("cast_9")];
+            tensor<fp16, [1]> clip_1_cast_fp16 = clip(alpha = const_384_to_fp16, beta = var_49_promoted_to_fp16, x = var_5547_to_fp16)[name = string("clip_1_cast_fp16")];
+            int32 one_hot_1_batch_dims_0 = const()[name = string("one_hot_1_batch_dims_0"), val = int32(0)];
+            bool one_hot_1_validate_indices_0 = const()[name = string("one_hot_1_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [128, 128]> to_onehot_identity_table_to_fp16 = const()[name = string("to_onehot_identity_table_to_fp16"), val = tensor<fp16, [128, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(556379072)))];
+            string prompt_id_to_int16_dtype_0 = const()[name = string("prompt_id_to_int16_dtype_0"), val = string("int16")];
+            string cast_230_dtype_0 = const()[name = string("cast_230_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
+            tensor<int16, [1]> prompt_id_to_int16 = cast(dtype = prompt_id_to_int16_dtype_0, x = prompt_id)[name = string("cast_8")];
+            tensor<int32, [1]> cast_230 = cast(dtype = cast_230_dtype_0, x = prompt_id_to_int16)[name = string("cast_7")];
+            tensor<bool, [1]> greater_equal_0 = greater_equal(x = cast_230, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
+            int32 slice_by_index_2 = const()[name = string("slice_by_index_2"), val = int32(128)];
+            tensor<int32, [1]> add_0 = add(x = cast_230, y = slice_by_index_2)[name = string("add_0")];
+            tensor<int32, [1]> select_0 = select(a = cast_230, b = add_0, cond = greater_equal_0)[name = string("select_0")];
+            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
+            string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
+            tensor<int16, [1]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_6")];
+            tensor<int32, [1]> cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_5")];
+            tensor<bool, [1]> greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
+            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(128)];
+            tensor<int32, [1]> add_0_1 = add(x = cast_0, y = slice_by_index_0)[name = string("add_0_1")];
+            tensor<int32, [1]> select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")];
+            int32 greater_equal_0_y_0_2 = const()[name = string("greater_equal_0_y_0_2"), val = int32(0)];
+            tensor<bool, [1]> greater_equal_0_2 = greater_equal(x = select_0_1, y = greater_equal_0_y_0_2)[name = string("greater_equal_0_2")];
+            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(128)];
+            tensor<int32, [1]> add_0_2 = add(x = select_0_1, y = slice_by_index_0_1)[name = string("add_0_2")];
+            tensor<int32, [1]> select_0_2 = select(a = select_0_1, b = add_0_2, cond = greater_equal_0_2)[name = string("select_0_2")];
+            int32 one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(0)];
+            tensor<fp16, [1, 128]> one_hot_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = one_hot_1_batch_dims_0, indices = select_0_2, validate_indices = one_hot_1_validate_indices_0, x = to_onehot_identity_table_to_fp16)[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16")];
+            tensor<int32, [1]> var_5593_axes_0 = const()[name = string("op_5593_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 128]> var_5593_cast_fp16 = expand_dims(axes = var_5593_axes_0, x = one_hot_1_cast_fp16_cast_uint16_cast_uint16)[name = string("op_5593_cast_fp16")];
+            tensor<int32, [3]> one_hot_reps_0 = const()[name = string("one_hot_reps_0"), val = tensor<int32, [3]>([1, 28, 1])];
+            tensor<fp16, [1, 28, 128]> one_hot_cast_fp16 = tile(reps = one_hot_reps_0, x = var_5593_cast_fp16)[name = string("one_hot_cast_fp16")];
+            int32 var_5602 = const()[name = string("op_5602"), val = int32(-1)];
+            bool input_1273_interleave_0 = const()[name = string("input_1273_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 28, 1152]> input_1273_cast_fp16 = concat(axis = var_5602, interleave = input_1273_interleave_0, values = (audio_signal_cast_fp16, one_hot_cast_fp16))[name = string("input_1273_cast_fp16")];
+            tensor<fp16, [2048, 1152]> prompt_kernel_0_weight_to_fp16 = const()[name = string("prompt_kernel_0_weight_to_fp16"), val = tensor<fp16, [2048, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(556411904)))];
+            tensor<fp16, [2048]> prompt_kernel_0_bias_to_fp16 = const()[name = string("prompt_kernel_0_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(561130560)))];
+            tensor<fp16, [1, 28, 2048]> linear_217_cast_fp16 = linear(bias = prompt_kernel_0_bias_to_fp16, weight = prompt_kernel_0_weight_to_fp16, x = input_1273_cast_fp16)[name = string("linear_217_cast_fp16")];
+            tensor<fp16, [1, 28, 2048]> input_cast_fp16 = relu(x = linear_217_cast_fp16)[name = string("input_cast_fp16")];
+            tensor<fp16, [1024, 2048]> prompt_kernel_2_weight_to_fp16 = const()[name = string("prompt_kernel_2_weight_to_fp16"), val = tensor<fp16, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(561134720)))];
+            tensor<fp16, [1024]> prompt_kernel_2_bias_to_fp16 = const()[name = string("prompt_kernel_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565329088)))];
+            tensor<fp16, [1, 28, 1024]> linear_218_cast_fp16 = linear(bias = prompt_kernel_2_bias_to_fp16, weight = prompt_kernel_2_weight_to_fp16, x = input_cast_fp16)[name = string("linear_218_cast_fp16")];
+            tensor<int32, [3]> var_5615_perm_0 = const()[name = string("op_5615_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string var_5615_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5615_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            string var_5620_dtype_0 = const()[name = string("op_5620_dtype_0"), val = string("int32")];
+            tensor<int32, [4]> var_5623_perm_0 = const()[name = string("op_5623_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string var_5623_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5623_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [4]> var_5626_perm_0 = const()[name = string("op_5626_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string var_5626_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5626_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            string var_5631_dtype_0 = const()[name = string("op_5631_dtype_0"), val = string("int32")];
+            tensor<int32, [1]> cache_len_out = cast(dtype = var_5631_dtype_0, x = clip_1_cast_fp16)[name = string("cast_0")];
+            tensor<fp16, [1, 24, 1024, 8]> var_5626_cast_fp16 = transpose(perm = var_5626_perm_0, x = obj_7_cast_fp16)[name = string("transpose_144")];
+            tensor<fp32, [1, 24, 1024, 8]> cache_time_out = cast(dtype = var_5626_cast_fp16_to_fp32_dtype_0, x = var_5626_cast_fp16)[name = string("cast_1")];
+            tensor<fp16, [1, 24, 42, 1024]> var_5623_cast_fp16 = transpose(perm = var_5623_perm_0, x = obj_5_cast_fp16)[name = string("transpose_145")];
+            tensor<fp32, [1, 24, 42, 1024]> cache_channel_out = cast(dtype = var_5623_cast_fp16_to_fp32_dtype_0, x = var_5623_cast_fp16)[name = string("cast_2")];
+            tensor<int32, [1]> encoded_length = cast(dtype = var_5620_dtype_0, x = clip_0_cast_fp16)[name = string("cast_3")];
+            tensor<fp16, [1, 1024, 28]> var_5615_cast_fp16 = transpose(perm = var_5615_perm_0, x = linear_218_cast_fp16)[name = string("transpose_146")];
+            tensor<fp32, [1, 1024, 28]> encoded = cast(dtype = var_5615_cast_fp16_to_fp32_dtype_0, x = var_5615_cast_fp16)[name = string("cast_4")];
+        } -> (encoded, encoded_length, cache_channel_out, cache_time_out, cache_len_out);
+}
\ No newline at end of file
diff --git a/fr/2240ms/encoder.mlmodelc/weights/weight.bin b/fr/2240ms/encoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..7a12bc272fe95f2080e5cebe63d7189329748a69
--- /dev/null
+++ b/fr/2240ms/encoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:728cf78b85d85c573a2a33a8287ea44bc7a846fc8462d9038d1ed1b24a2c9ac8
+size 565331200
diff --git a/fr/2240ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/2240ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..808c1ca295fe918d2a2372f1e2fadb3bc18930d1
--- /dev/null
+++ b/fr/2240ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a0653e7c223390a9fe23c5e6f6266b97488e294bbc759705aaa66c68bb15ff8f
+size 805589
diff --git a/fr/2240ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/2240ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..7a12bc272fe95f2080e5cebe63d7189329748a69
--- /dev/null
+++ b/fr/2240ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:728cf78b85d85c573a2a33a8287ea44bc7a846fc8462d9038d1ed1b24a2c9ac8
+size 565331200
diff --git a/fr/2240ms/encoder.mlpackage/Manifest.json b/fr/2240ms/encoder.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..955535c4db53bb00e2d26147a70036cd887e0451
--- /dev/null
+++ b/fr/2240ms/encoder.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "00BBBFAC-20B6-427A-B978-654939D77446": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "5707DA36-A222-47FA-8727-5AED43CBE002": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "5707DA36-A222-47FA-8727-5AED43CBE002"
+}
diff --git a/fr/2240ms/joint.mlmodelc/analytics/coremldata.bin b/fr/2240ms/joint.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..38aade026cc9376471863f9f09be292c9fe8578b
--- /dev/null
+++ b/fr/2240ms/joint.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad1f6f5571ee9e7ad8a533b4679ed7416879f4d8e5fc10f40835c1d014b822a4
+size 243
diff --git a/fr/2240ms/joint.mlmodelc/coremldata.bin b/fr/2240ms/joint.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..45df627e7953068f36ea8f3e5467664cf318a44e
--- /dev/null
+++ b/fr/2240ms/joint.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e630dfefcd0f652f2815f76cfbacb89b51f2c94618fbc517e41779426099a1f
+size 341
diff --git a/fr/2240ms/joint.mlmodelc/model.mil b/fr/2240ms/joint.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..041dd87f66df1673f6226d5c642c93a64d4e2730
--- /dev/null
+++ b/fr/2240ms/joint.mlmodelc/model.mil
@@ -0,0 +1,31 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp32, [1, 640, 1]> decoder, tensor<fp32, [1, 1024, 1]> encoder) {
+            tensor<int32, [3]> input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [640, 1024]> module_enc_weight_to_fp16 = const()[name = string("module_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [640]> module_enc_bias_to_fp16 = const()[name = string("module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1310848)))];
+            tensor<fp16, [1, 1024, 1]> encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_2")];
+            tensor<fp16, [1, 1, 1024]> input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_to_fp16)[name = string("transpose_1")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = module_enc_bias_to_fp16, weight = module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> module_pred_weight_to_fp16 = const()[name = string("module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1312192)))];
+            tensor<fp16, [640]> module_pred_bias_to_fp16 = const()[name = string("module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2131456)))];
+            tensor<fp16, [1, 640, 1]> decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_1")];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_to_fp16)[name = string("transpose_0")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = module_pred_bias_to_fp16, weight = module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_23_axes_0 = const()[name = string("op_23_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = string("op_23_cast_fp16")];
+            tensor<int32, [1]> var_25_axes_0 = const()[name = string("op_25_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_25_cast_fp16 = expand_dims(axes = var_25_axes_0, x = linear_1_cast_fp16)[name = string("op_25_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_25_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("input_7_cast_fp16")];
+            tensor<fp16, [849, 640]> module_joint_net_2_weight_to_fp16 = const()[name = string("module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2132800)))];
+            tensor<fp16, [849]> module_joint_net_2_bias_to_fp16 = const()[name = string("module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [849]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3219584)))];
+            tensor<fp16, [1, 1, 1, 849]> linear_2_cast_fp16 = linear(bias = module_joint_net_2_bias_to_fp16, weight = module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_2_cast_fp16")];
+            string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 1, 1, 849]> logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_0")];
+        } -> (logits);
+}
\ No newline at end of file
diff --git a/fr/2240ms/joint.mlmodelc/weights/weight.bin b/fr/2240ms/joint.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..109a5bf4eb27d11fcd040ea350a8e6d00d18932a
--- /dev/null
+++ b/fr/2240ms/joint.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9fa5fb0d14d87484d9ffd1f4d30996ea456c14ea3e5ab9d910486d8e86d3a540
+size 3221346
diff --git a/fr/2240ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/2240ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..0c650fe487d18d0d937a1fc295a3eee53b4aacbe
--- /dev/null
+++ b/fr/2240ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4bac39a8a21b3e7e8cdeba1d03b92e8f743bf87cb2de3b1d85c505a580a0b218
+size 4486
diff --git a/fr/2240ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/2240ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..109a5bf4eb27d11fcd040ea350a8e6d00d18932a
--- /dev/null
+++ b/fr/2240ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9fa5fb0d14d87484d9ffd1f4d30996ea456c14ea3e5ab9d910486d8e86d3a540
+size 3221346
diff --git a/fr/2240ms/joint.mlpackage/Manifest.json b/fr/2240ms/joint.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7e49a7390ba8e327aff397061f0fd4476b7c70a
--- /dev/null
+++ b/fr/2240ms/joint.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "0DC18C6B-B3DE-4D3A-80EE-28CAA06F6708": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "88105E4A-1C49-40F4-A835-C9B725CEEC59": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "88105E4A-1C49-40F4-A835-C9B725CEEC59"
+}
diff --git a/fr/2240ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin b/fr/2240ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..a856ad184cad37664eb30bb699b3a88e6401000d
--- /dev/null
+++ b/fr/2240ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28b5c13f8ce2d8154edb31968c174d322ee4ae76db42f41ecf7ac5e0fd87f57d
+size 243
diff --git a/fr/2240ms/joint_noencproj_batched.mlmodelc/coremldata.bin b/fr/2240ms/joint_noencproj_batched.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..b68dc2ebdbe0b8957cd45ed4b265fda6b6195204
--- /dev/null
+++ b/fr/2240ms/joint_noencproj_batched.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5dc74778713432d34163ae80901d3294eef3341dd020a29dc05d67eba8f6f0a
+size 406
diff --git a/fr/2240ms/joint_noencproj_batched.mlmodelc/model.mil b/fr/2240ms/joint_noencproj_batched.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..b66479ae4eaa92652a98a34a18e993259a14e6c0
--- /dev/null
+++ b/fr/2240ms/joint_noencproj_batched.mlmodelc/model.mil
@@ -0,0 +1,26 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios18>(tensor<fp32, [1, 640, 1]> decoder, tensor<fp32, [1, 4, 640]> encoder_proj) {
+            tensor<int32, [3]> input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819328)))];
+            tensor<fp16, [1, 640, 1]> decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_2")];
+            tensor<fp16, [1, 1, 640]> input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = decoder_to_fp16)[name = string("transpose_0")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [1]> var_15_axes_0 = const()[name = string("op_15_axes_0"), val = tensor<int32, [1]>([2])];
+            string encoder_proj_to_fp16_dtype_0 = const()[name = string("encoder_proj_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 4, 640]> encoder_proj_to_fp16 = cast(dtype = encoder_proj_to_fp16_dtype_0, x = encoder_proj)[name = string("cast_1")];
+            tensor<fp16, [1, 4, 1, 640]> var_15_cast_fp16 = expand_dims(axes = var_15_axes_0, x = encoder_proj_to_fp16)[name = string("op_15_cast_fp16")];
+            tensor<int32, [1]> var_17_axes_0 = const()[name = string("op_17_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_17_cast_fp16 = expand_dims(axes = var_17_axes_0, x = linear_0_cast_fp16)[name = string("op_17_cast_fp16")];
+            tensor<fp16, [1, 4, 1, 640]> input_3_cast_fp16 = add(x = var_15_cast_fp16, y = var_17_cast_fp16)[name = string("input_3_cast_fp16")];
+            tensor<fp16, [1, 4, 1, 640]> input_5_cast_fp16 = relu(x = input_3_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<fp16, [849, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(820672)))];
+            tensor<fp16, [849]> joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [849]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1907456)))];
+            tensor<fp16, [1, 4, 1, 849]> linear_1_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_5_cast_fp16)[name = string("linear_1_cast_fp16")];
+            string linear_1_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_1_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 4, 1, 849]> logits = cast(dtype = linear_1_cast_fp16_to_fp32_dtype_0, x = linear_1_cast_fp16)[name = string("cast_0")];
+        } -> (logits);
+}
\ No newline at end of file
diff --git a/fr/2240ms/joint_noencproj_batched.mlmodelc/weights/weight.bin b/fr/2240ms/joint_noencproj_batched.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..4e66049a9a8bde2b92859df3192b671ddf9ab81d
--- /dev/null
+++ b/fr/2240ms/joint_noencproj_batched.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9a9dd24e6b63b97540b0db0e990418bc6fddc9c2bf85b82eb226e6df1ea02a4
+size 1909218
diff --git a/fr/2240ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/2240ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..4f82db74749dc3350042c3f6c1453c4d4e620d50
--- /dev/null
+++ b/fr/2240ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d13f08c4961187e47001773d7d4d5d5937ba7ff68475eb7c1bac41d7014c64a1
+size 3839
diff --git a/fr/2240ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/2240ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..4e66049a9a8bde2b92859df3192b671ddf9ab81d
--- /dev/null
+++ b/fr/2240ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9a9dd24e6b63b97540b0db0e990418bc6fddc9c2bf85b82eb226e6df1ea02a4
+size 1909218
diff --git a/fr/2240ms/joint_noencproj_batched.mlpackage/Manifest.json b/fr/2240ms/joint_noencproj_batched.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..5f2e3080d9f6dc5617b9630128af453e31562411
--- /dev/null
+++ b/fr/2240ms/joint_noencproj_batched.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "EFDDA9F1-A088-4CE2-8A64-5928EB8F429C": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "F28FF7B3-AC0A-4ACA-93F8-DF201DFFE75B": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "F28FF7B3-AC0A-4ACA-93F8-DF201DFFE75B"
+}
diff --git a/fr/2240ms/metadata.json b/fr/2240ms/metadata.json
new file mode 100644
index 0000000000000000000000000000000000000000..29f761ec59fec1e324129193a400fe123c84bc21
--- /dev/null
+++ b/fr/2240ms/metadata.json
@@ -0,0 +1,199 @@
+{
+  "model": "nvidia/nemotron-asr-streaming-multilingual-0.6b",
+  "model_class": "nemo.collections.asr.models.rnnt_bpe_models_prompt.EncDecRNNTBPEModelWithPrompt",
+  "sample_rate": 16000,
+  "mel_features": 128,
+  "chunk_mel_frames": 224,
+  "pre_encode_cache": 9,
+  "total_mel_frames": 233,
+  "att_context_size": [
+    42,
+    13
+  ],
+  "vocab_size": 848,
+  "blank_idx": 848,
+  "vocab_pruned": true,
+  "vocab_pruned_original_size": 13087,
+  "cache_channel_shape": [
+    1,
+    24,
+    42,
+    1024
+  ],
+  "cache_time_shape": [
+    1,
+    24,
+    1024,
+    8
+  ],
+  "decoder_hidden": 640,
+  "decoder_layers": 2,
+  "encoder_dim": 1024,
+  "num_prompts": 128,
+  "prompt_dictionary": {
+    "af-ZA": 54,
+    "am-ET": 49,
+    "ar": 7,
+    "ar-AR": 7,
+    "auto": 101,
+    "ay-BO": 81,
+    "az-AZ": 66,
+    "bg": 30,
+    "bg-BG": 30,
+    "bn-IN": 36,
+    "cs": 22,
+    "cs-CZ": 22,
+    "da": 25,
+    "da-DK": 25,
+    "de": 9,
+    "de-DE": 9,
+    "el": 21,
+    "el-GR": 21,
+    "en": 0,
+    "en-GB": 1,
+    "en-US": 0,
+    "enGB": 1,
+    "es": 3,
+    "es-ES": 2,
+    "es-US": 3,
+    "esES": 2,
+    "et": 60,
+    "et-EE": 60,
+    "fa-IR": 38,
+    "fi": 26,
+    "fi-FI": 26,
+    "fr": 8,
+    "fr-CA": 100,
+    "fr-FR": 8,
+    "gn-PY": 82,
+    "gu-IN": 42,
+    "ha-NG": 50,
+    "haw-US": 97,
+    "he-IL": 64,
+    "hi": 6,
+    "hi-HI": 6,
+    "hi-IN": 6,
+    "hr": 29,
+    "hr-HR": 29,
+    "hu": 23,
+    "hu-HU": 23,
+    "hy-AM": 68,
+    "id-ID": 34,
+    "ig-NG": 53,
+    "it": 15,
+    "it-IT": 15,
+    "ja-JA": 10,
+    "ja-JP": 10,
+    "ka-GE": 67,
+    "km-KH": 47,
+    "kn-IN": 43,
+    "ko": 14,
+    "ko-KO": 14,
+    "ko-KR": 14,
+    "ku-TR": 65,
+    "ky-KG": 71,
+    "ln-CD": 58,
+    "lt": 31,
+    "lt-LT": 31,
+    "lv": 61,
+    "lv-LV": 61,
+    "mi-NZ": 96,
+    "ml-IN": 44,
+    "mr-IN": 41,
+    "ms-MY": 35,
+    "mt-MT": 102,
+    "nah-MX": 83,
+    "nb": 103,
+    "nb-NO": 103,
+    "ne-NP": 46,
+    "nl": 16,
+    "nl-NL": 16,
+    "nn": 104,
+    "nn-NO": 104,
+    "no": 27,
+    "no-NO": 27,
+    "ny-MW": 57,
+    "or-KE": 59,
+    "pl": 17,
+    "pl-PL": 17,
+    "pt": 13,
+    "pt-BR": 12,
+    "pt-PT": 13,
+    "qu-PE": 80,
+    "ro": 20,
+    "ro-RO": 20,
+    "ru": 11,
+    "ru-RU": 11,
+    "rw-RW": 55,
+    "si-LK": 45,
+    "sk": 28,
+    "sk-SK": 28,
+    "sl": 62,
+    "sl-SI": 62,
+    "sm-WS": 98,
+    "so-SO": 56,
+    "sv": 24,
+    "sv-SE": 24,
+    "sw-KE": 48,
+    "ta-IN": 39,
+    "te-IN": 40,
+    "tg-TJ": 70,
+    "th-TH": 32,
+    "to-TO": 99,
+    "tr": 18,
+    "tr-TR": 18,
+    "uk": 19,
+    "uk-UA": 19,
+    "ur-PK": 37,
+    "uz-UZ": 69,
+    "vi-VN": 33,
+    "yo-NG": 52,
+    "zh-CN": 4,
+    "zh-TW": 5,
+    "zh-ZH": 4,
+    "zu-ZA": 51
+  },
+  "default_prompt_id": 101,
+  "lang_tag_token_ids": [
+    1,
+    72,
+    122,
+    170,
+    230,
+    231,
+    263,
+    282,
+    390,
+    415,
+    430,
+    483,
+    501,
+    519,
+    542,
+    565,
+    594,
+    619,
+    620,
+    629,
+    640,
+    659,
+    660,
+    677,
+    704,
+    717,
+    735,
+    741,
+    790,
+    791,
+    793,
+    794,
+    795,
+    800,
+    819,
+    829,
+    830,
+    844,
+    845
+  ],
+  "chunk_ms": 2240
+}
\ No newline at end of file
diff --git a/fr/2240ms/preprocessor.mlmodelc/analytics/coremldata.bin b/fr/2240ms/preprocessor.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..d4b1561ab413a9d87db506bc842f077779dcbded
--- /dev/null
+++ b/fr/2240ms/preprocessor.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b8033bec5ee01649f325b8f4c5aeef1b31c99b469ce56d46039c1b73f09585d
+size 243
diff --git a/fr/2240ms/preprocessor.mlmodelc/coremldata.bin b/fr/2240ms/preprocessor.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..50c6f5fdff7aca33e5a111a3ad36751334f6bcbc
--- /dev/null
+++ b/fr/2240ms/preprocessor.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da7084cfdd3b4e9c7c91f3fac1a37afcce7ae905e8b799e72f99aa4fdf6664ab
+size 431
diff --git a/fr/2240ms/preprocessor.mlmodelc/model.mil b/fr/2240ms/preprocessor.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..b1a0b2b9193c992de42e51245fc1ef433d345afc
--- /dev/null
+++ b/fr/2240ms/preprocessor.mlmodelc/model.mil
@@ -0,0 +1,122 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios18>(tensor<fp32, [1, ?]> audio, tensor<int32, [1]> audio_length) [FlexibleShapeInformation = tuple<tuple<string, dict<string, tensor<int32, [?]>>>, tuple<string, dict<string, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"audio", [1, 1]}}), ("RangeDims", {{"audio", [[1, 1], [1, 1280000]]}})))] {
+            int32 var_9 = const()[name = string("op_9"), val = int32(1)];
+            int32 var_10 = const()[name = string("op_10"), val = int32(160)];
+            int32 var_12 = const()[name = string("op_12"), val = int32(0)];
+            int32 var_33 = const()[name = string("op_33"), val = int32(512)];
+            tensor<int32, [1]> var_34 = add(x = audio_length, y = var_33)[name = string("op_34")];
+            int32 var_35 = const()[name = string("op_35"), val = int32(512)];
+            tensor<int32, [1]> var_36 = sub(x = var_34, y = var_35)[name = string("op_36")];
+            tensor<int32, [1]> floor_div_0 = floor_div(x = var_36, y = var_10)[name = string("floor_div_0")];
+            tensor<bool, [1]> var_39 = equal(x = audio_length, y = var_12)[name = string("op_39")];
+            tensor<int32, [1]> var_40 = const()[name = string("op_40"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> mel_length = select(a = var_40, b = floor_div_0, cond = var_39)[name = string("seq_len")];
+            string audio_to_fp16_dtype_0 = const()[name = string("audio_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, ?]> audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = string("cast_10")];
+            tensor<int32, [2]> var_42_shape_cast_fp16 = shape(x = audio_to_fp16)[name = string("op_42_shape_cast_fp16")];
+            int32 gather_0_axis_0 = const()[name = string("gather_0_axis_0"), val = int32(0)];
+            int32 gather_0_batch_dims_0 = const()[name = string("gather_0_batch_dims_0"), val = int32(0)];
+            bool gather_0_validate_indices_0 = const()[name = string("gather_0_validate_indices_0"), val = bool(false)];
+            string var_42_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_42_shape_cast_fp16_to_int16_dtype_0"), val = string("int16")];
+            uint16 gather_0_indices_0_to_uint16 = const()[name = string("gather_0_indices_0_to_uint16"), val = uint16(1)];
+            tensor<int16, [2]> var_42_shape_cast_fp16_to_int16 = cast(dtype = var_42_shape_cast_fp16_to_int16_dtype_0, x = var_42_shape_cast_fp16)[name = string("cast_9")];
+            int16 gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = gather_0_indices_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_42_shape_cast_fp16_to_int16)[name = string("gather_0_cast_uint16")];
+            string gather_0_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_0_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            int32 const_0 = const()[name = string("const_0"), val = int32(0)];
+            int32 const_1 = const()[name = string("const_1"), val = int32(1)];
+            int32 gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = string("cast_8")];
+            tensor<int32, [?]> var_43 = range_1d(end = gather_0_cast_uint16_to_int32, start = const_0, step = const_1)[name = string("op_43")];
+            tensor<int32, [1]> var_44_axes_0 = const()[name = string("op_44_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> var_44 = expand_dims(axes = var_44_axes_0, x = var_43)[name = string("op_44")];
+            tensor<int32, [1]> var_45_axes_0 = const()[name = string("op_45_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_45 = expand_dims(axes = var_45_axes_0, x = audio_length)[name = string("op_45")];
+            tensor<bool, [1, ?]> timemask = less(x = var_44, y = var_45)[name = string("timemask")];
+            tensor<int32, [2]> var_48_begin_0 = const()[name = string("op_48_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_48_end_0 = const()[name = string("op_48_end_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<bool, [2]> var_48_end_mask_0 = const()[name = string("op_48_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<bool, [2]> var_48_squeeze_mask_0 = const()[name = string("op_48_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
+            tensor<fp16, [1]> var_48_cast_fp16 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, squeeze_mask = var_48_squeeze_mask_0, x = audio_to_fp16)[name = string("op_48_cast_fp16")];
+            tensor<int32, [1]> var_49_axes_0 = const()[name = string("op_49_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1]> var_49_cast_fp16 = expand_dims(axes = var_49_axes_0, x = var_48_cast_fp16)[name = string("op_49_cast_fp16")];
+            tensor<int32, [2]> var_51_begin_0 = const()[name = string("op_51_begin_0"), val = tensor<int32, [2]>([0, 1])];
+            tensor<int32, [2]> var_51_end_0 = const()[name = string("op_51_end_0"), val = tensor<int32, [2]>([1, 0])];
+            tensor<bool, [2]> var_51_end_mask_0 = const()[name = string("op_51_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<fp16, [1, ?]> var_51_cast_fp16 = slice_by_index(begin = var_51_begin_0, end = var_51_end_0, end_mask = var_51_end_mask_0, x = audio_to_fp16)[name = string("op_51_cast_fp16")];
+            tensor<int32, [2]> var_53_begin_0 = const()[name = string("op_53_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_53_end_0 = const()[name = string("op_53_end_0"), val = tensor<int32, [2]>([1, -1])];
+            tensor<bool, [2]> var_53_end_mask_0 = const()[name = string("op_53_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp16, [1, ?]> var_53_cast_fp16 = slice_by_index(begin = var_53_begin_0, end = var_53_end_0, end_mask = var_53_end_mask_0, x = audio_to_fp16)[name = string("op_53_cast_fp16")];
+            fp16 var_54_to_fp16 = const()[name = string("op_54_to_fp16"), val = fp16(0x1.f0cp-1)];
+            tensor<fp16, [1, ?]> var_55_cast_fp16 = mul(x = var_53_cast_fp16, y = var_54_to_fp16)[name = string("op_55_cast_fp16")];
+            tensor<fp16, [1, ?]> var_56_cast_fp16 = sub(x = var_51_cast_fp16, y = var_55_cast_fp16)[name = string("op_56_cast_fp16")];
+            bool x_3_interleave_0 = const()[name = string("x_3_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, ?]> x_3_cast_fp16 = concat(axis = var_9, interleave = x_3_interleave_0, values = (var_49_cast_fp16, var_56_cast_fp16))[name = string("x_3_cast_fp16")];
+            tensor<bool, [1, ?]> var_59 = logical_not(x = timemask)[name = string("op_59")];
+            fp16 var_16_to_fp16 = const()[name = string("op_16_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, ?]> input_1_cast_fp16 = select(a = var_16_to_fp16, b = x_3_cast_fp16, cond = var_59)[name = string("input_1_cast_fp16")];
+            tensor<int32, [3]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, ?]> input_3_cast_fp16 = reshape(shape = concat_1x, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")];
+            tensor<int32, [6]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 256, 256])];
+            string input_5_mode_0 = const()[name = string("input_5_mode_0"), val = string("constant")];
+            fp16 const_3_to_fp16 = const()[name = string("const_3_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, ?]> input_5_cast_fp16 = pad(constant_val = const_3_to_fp16, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<int32, [2]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [2]>([1, -1])];
+            tensor<fp16, [1, ?]> input_cast_fp16 = reshape(shape = concat_2x, x = input_5_cast_fp16)[name = string("input_cast_fp16")];
+            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([160])];
+            tensor<int32, [1]> expand_dims_4_axes_0 = const()[name = string("expand_dims_4_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, ?]> expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = input_cast_fp16)[name = string("expand_dims_4_cast_fp16")];
+            string conv_0_pad_type_0 = const()[name = string("conv_0_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> conv_0_pad_0 = const()[name = string("conv_0_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_0_dilations_0 = const()[name = string("conv_0_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 conv_0_groups_0 = const()[name = string("conv_0_groups_0"), val = int32(1)];
+            tensor<fp16, [257, 1, 512]> expand_dims_1_to_fp16 = const()[name = string("expand_dims_1_to_fp16"), val = tensor<fp16, [257, 1, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 257, ?]> conv_0_cast_fp16 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_0_cast_fp16")];
+            string conv_1_pad_type_0 = const()[name = string("conv_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> conv_1_pad_0 = const()[name = string("conv_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_1_dilations_0 = const()[name = string("conv_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 conv_1_groups_0 = const()[name = string("conv_1_groups_0"), val = int32(1)];
+            tensor<fp16, [257, 1, 512]> expand_dims_2_to_fp16 = const()[name = string("expand_dims_2_to_fp16"), val = tensor<fp16, [257, 1, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263296)))];
+            tensor<fp16, [1, 257, ?]> conv_1_cast_fp16 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_1_cast_fp16")];
+            int32 stack_0_axis_0 = const()[name = string("stack_0_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 257, ?, 2]> stack_0_cast_fp16 = stack(axis = stack_0_axis_0, values = (conv_0_cast_fp16, conv_1_cast_fp16))[name = string("stack_0_cast_fp16")];
+            fp16 var_19_promoted_to_fp16 = const()[name = string("op_19_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1, 257, ?, 2]> var_74_cast_fp16 = pow(x = stack_0_cast_fp16, y = var_19_promoted_to_fp16)[name = string("op_74_cast_fp16")];
+            tensor<int32, [1]> var_76_axes_0 = const()[name = string("op_76_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_76_keep_dims_0 = const()[name = string("op_76_keep_dims_0"), val = bool(false)];
+            tensor<fp16, [1, 257, ?]> var_76_cast_fp16 = reduce_sum(axes = var_76_axes_0, keep_dims = var_76_keep_dims_0, x = var_74_cast_fp16)[name = string("op_76_cast_fp16")];
+            tensor<fp16, [1, 257, ?]> x_11_cast_fp16 = identity(x = var_76_cast_fp16)[name = string("x_11_cast_fp16")];
+            bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)];
+            bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 128, 257]> const_4_to_fp16 = const()[name = string("const_4_to_fp16"), val = tensor<fp16, [1, 128, 257]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526528)))];
+            tensor<fp16, [1, 128, ?]> x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = const_4_to_fp16, y = x_11_cast_fp16)[name = string("x_13_cast_fp16")];
+            fp16 var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 128, ?]> var_84_cast_fp16 = add(x = x_13_cast_fp16, y = var_83_to_fp16)[name = string("op_84_cast_fp16")];
+            fp32 x_epsilon_0 = const()[name = string("x_epsilon_0"), val = fp32(0x1p-149)];
+            tensor<fp16, [1, 128, ?]> x_cast_fp16 = log(epsilon = x_epsilon_0, x = var_84_cast_fp16)[name = string("x_cast_fp16")];
+            tensor<int32, [3]> var_86_shape_cast_fp16 = shape(x = x_cast_fp16)[name = string("op_86_shape_cast_fp16")];
+            int32 gather_5_batch_dims_0 = const()[name = string("gather_5_batch_dims_0"), val = int32(0)];
+            bool gather_5_validate_indices_0 = const()[name = string("gather_5_validate_indices_0"), val = bool(false)];
+            string var_86_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_86_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
+            int32 gather_5_cast_uint16_axis_0 = const()[name = string("gather_5_cast_uint16_axis_0"), val = int32(0)];
+            uint16 select_0_to_uint16 = const()[name = string("select_0_to_uint16"), val = uint16(2)];
+            tensor<uint16, [3]> var_86_shape_cast_fp16_to_uint16 = cast(dtype = var_86_shape_cast_fp16_to_uint16_dtype_0, x = var_86_shape_cast_fp16)[name = string("cast_7")];
+            uint16 gather_5_cast_uint16_cast_uint16 = gather(axis = gather_5_cast_uint16_axis_0, batch_dims = gather_5_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_5_validate_indices_0, x = var_86_shape_cast_fp16_to_uint16)[name = string("gather_5_cast_uint16_cast_uint16")];
+            string gather_5_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_5_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            int32 const_5 = const()[name = string("const_5"), val = int32(0)];
+            int32 const_6 = const()[name = string("const_6"), val = int32(1)];
+            int32 gather_5_cast_uint16_to_int32 = cast(dtype = gather_5_cast_uint16_to_int32_dtype_0, x = gather_5_cast_uint16_cast_uint16)[name = string("cast_6")];
+            tensor<int32, [?]> mask_1 = range_1d(end = gather_5_cast_uint16_to_int32, start = const_5, step = const_6)[name = string("mask_1")];
+            tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = string("expand_dims_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = string("expand_dims_0")];
+            tensor<int32, [1]> var_91_axes_0 = const()[name = string("op_91_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_91 = expand_dims(axes = var_91_axes_0, x = mel_length)[name = string("op_91")];
+            tensor<bool, [1, ?]> mask = greater_equal(x = expand_dims_0, y = var_91)[name = string("mask")];
+            tensor<int32, [1]> var_93_axes_0 = const()[name = string("op_93_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, ?]> var_93 = expand_dims(axes = var_93_axes_0, x = mask)[name = string("op_93")];
+            tensor<fp16, [1, 128, ?]> processed_signal_cast_fp16 = select(a = var_16_to_fp16, b = x_cast_fp16, cond = var_93)[name = string("processed_signal_cast_fp16")];
+            string processed_signal_cast_fp16_to_fp32_dtype_0 = const()[name = string("processed_signal_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 128, ?]> mel = cast(dtype = processed_signal_cast_fp16_to_fp32_dtype_0, x = processed_signal_cast_fp16)[name = string("cast_5")];
+        } -> (mel, mel_length);
+}
\ No newline at end of file
diff --git a/fr/2240ms/preprocessor.mlmodelc/weights/weight.bin b/fr/2240ms/preprocessor.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411
--- /dev/null
+++ b/fr/2240ms/preprocessor.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3
+size 592384
diff --git a/fr/2240ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/2240ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..050fa97ca7a2aa4b7c4fa318f4fa2a51914287c4
--- /dev/null
+++ b/fr/2240ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0feb90f2cf9a5bc749fa763b6eb78a23755b6184b41a262fe08faebe4e709b3e
+size 16035
diff --git a/fr/2240ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/2240ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411
--- /dev/null
+++ b/fr/2240ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3
+size 592384
diff --git a/fr/2240ms/preprocessor.mlpackage/Manifest.json b/fr/2240ms/preprocessor.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..667b90978564849cb1c6d3c8b10f535ad524d4be
--- /dev/null
+++ b/fr/2240ms/preprocessor.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "02ED6D10-6658-4E52-83B8-E0D3DEA2B8AC": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "8D115BDF-817A-4C02-9F24-4AF6137B0210": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "8D115BDF-817A-4C02-9F24-4AF6137B0210"
+}
diff --git a/fr/2240ms/tokenizer.json b/fr/2240ms/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..53c1afe82c2a72808c78fffb846345d590c453fd
--- /dev/null
+++ b/fr/2240ms/tokenizer.json
@@ -0,0 +1,851 @@
+{
+  "0": "<unk>",
+  "1": "<bg-BG>",
+  "2": "▁",
+  "3": ".",
+  "4": ",",
+  "5": "e",
+  "6": "t",
+  "7": "a",
+  "8": "s",
+  "9": "o",
+  "10": "i",
+  "11": "r",
+  "12": "l",
+  "13": "u",
+  "14": "d",
+  "15": "c",
+  "16": "h",
+  "17": "m",
+  "18": "p",
+  "19": "n",
+  "20": "g",
+  "21": "f",
+  "22": "en",
+  "23": "in",
+  "24": "on",
+  "25": "y",
+  "26": "er",
+  "27": "an",
+  "28": "w",
+  "29": "0",
+  "30": "b",
+  "31": "v",
+  "32": "2",
+  "33": "1",
+  "34": "▁the",
+  "35": "k",
+  "36": "3",
+  "37": "5",
+  "38": "q",
+  "39": "I",
+  "40": "é",
+  "41": "4",
+  "42": "z",
+  "43": "A",
+  "44": "6",
+  "45": "j",
+  "46": "E",
+  "47": "7",
+  "48": "T",
+  "49": "8",
+  "50": "9",
+  "51": "S",
+  "52": "x",
+  "53": "C",
+  "54": "M",
+  "55": "P",
+  "56": "D",
+  "57": "B",
+  "58": "U",
+  "59": "N",
+  "60": "O",
+  "61": "G",
+  "62": "à",
+  "63": "L",
+  "64": "è",
+  "65": "V",
+  "66": "R",
+  "67": "ó",
+  "68": "J",
+  "69": "K",
+  "70": "ê",
+  "71": "ç",
+  "72": "<cs-CZ>",
+  "73": "st",
+  "74": "ch",
+  "75": "le",
+  "76": "li",
+  "77": "▁po",
+  "78": "no",
+  "79": "to",
+  "80": "me",
+  "81": "te",
+  "82": "ho",
+  "83": "▁pro",
+  "84": "ro",
+  "85": "▁na",
+  "86": "ce",
+  "87": "la",
+  "88": "ni",
+  "89": "ra",
+  "90": "ti",
+  "91": "lo",
+  "92": "ko",
+  "93": "po",
+  "94": "je",
+  "95": "de",
+  "96": "na",
+  "97": "mi",
+  "98": "ci",
+  "99": "▁by",
+  "100": "ve",
+  "101": "▁za",
+  "102": "▁A",
+  "103": "re",
+  "104": "ou",
+  "105": "vo",
+  "106": "né",
+  "107": "va",
+  "108": "mo",
+  "109": "ze",
+  "110": "ne",
+  "111": "ka",
+  "112": "▁jak",
+  "113": "ny",
+  "114": "vi",
+  "115": "ent",
+  "116": "▁pan",
+  "117": "Z",
+  "118": "▁To",
+  "119": "▁Je",
+  "120": "ă",
+  "121": "X",
+  "122": "<da-DK>",
+  "123": "▁for",
+  "124": "▁at",
+  "125": "et",
+  "126": "▁vi",
+  "127": "al",
+  "128": "▁de",
+  "129": "▁der",
+  "130": "or",
+  "131": "om",
+  "132": "and",
+  "133": "at",
+  "134": "▁af",
+  "135": "ge",
+  "136": "ar",
+  "137": "is",
+  "138": "▁be",
+  "139": "un",
+  "140": "lig",
+  "141": "▁man",
+  "142": "ig",
+  "143": "▁som",
+  "144": "el",
+  "145": "ag",
+  "146": "erne",
+  "147": "▁den",
+  "148": "ste",
+  "149": "id",
+  "150": "ske",
+  "151": "iv",
+  "152": "▁her",
+  "153": "ion",
+  "154": "am",
+  "155": "ur",
+  "156": "for",
+  "157": "▁pr",
+  "158": "▁sig",
+  "159": "▁ind",
+  "160": "ende",
+  "161": "▁Vi",
+  "162": "ation",
+  "163": "mme",
+  "164": "tage",
+  "165": "▁op",
+  "166": "▁fra",
+  "167": "▁alle",
+  "168": "▁Men",
+  "169": "▁var",
+  "170": "<de-DE>",
+  "171": "▁die",
+  "172": "sch",
+  "173": "it",
+  "174": "gen",
+  "175": "▁W",
+  "176": "▁B",
+  "177": "▁E",
+  "178": "▁F",
+  "179": "ll",
+  "180": "▁es",
+  "181": "▁K",
+  "182": "ie",
+  "183": "au",
+  "184": "ich",
+  "185": "ck",
+  "186": "ten",
+  "187": "mal",
+  "188": "ein",
+  "189": "▁T",
+  "190": "▁mit",
+  "191": "hr",
+  "192": "ter",
+  "193": "tz",
+  "194": "▁G",
+  "195": "ben",
+  "196": "um",
+  "197": "us",
+  "198": "cht",
+  "199": "il",
+  "200": "ut",
+  "201": "▁ver",
+  "202": "▁hier",
+  "203": "▁von",
+  "204": "ri",
+  "205": "ach",
+  "206": "ol",
+  "207": "▁Da",
+  "208": "sp",
+  "209": "ell",
+  "210": "▁ja",
+  "211": "uch",
+  "212": "wi",
+  "213": "rei",
+  "214": "▁Ge",
+  "215": "und",
+  "216": "isch",
+  "217": "▁sie",
+  "218": "▁Ja",
+  "219": "▁du",
+  "220": "▁dem",
+  "221": "▁Sch",
+  "222": "▁Ma",
+  "223": "▁De",
+  "224": "▁Le",
+  "225": "▁In",
+  "226": "▁Re",
+  "227": "▁sein",
+  "228": "▁Mi",
+  "229": "▁Ha",
+  "230": "<el-GR>",
+  "231": "<et-EE>",
+  "232": "ma",
+  "233": "ta",
+  "234": "se",
+  "235": "da",
+  "236": "si",
+  "237": "ks",
+  "238": "ga",
+  "239": "he",
+  "240": "mu",
+  "241": "tu",
+  "242": "ha",
+  "243": "ja",
+  "244": "gi",
+  "245": "nd",
+  "246": "oo",
+  "247": "gu",
+  "248": "ju",
+  "249": "est",
+  "250": "▁pa",
+  "251": "tud",
+  "252": "▁pea",
+  "253": "sta",
+  "254": "di",
+  "255": "mise",
+  "256": "▁su",
+  "257": "ide",
+  "258": "val",
+  "259": "▁Me",
+  "260": "ment",
+  "261": "▁meil",
+  "262": "Q",
+  "263": "<fi-FI>",
+  "264": "ssa",
+  "265": "lla",
+  "266": "ty",
+  "267": "ki",
+  "268": "pa",
+  "269": "lle",
+  "270": "lu",
+  "271": "tta",
+  "272": "isi",
+  "273": "ise",
+  "274": "tte",
+  "275": "ista",
+  "276": "inen",
+  "277": "llis",
+  "278": "vu",
+  "279": "▁voi",
+  "280": "iden",
+  "281": "ë",
+  "282": "<fr-FR>",
+  "283": "▁est",
+  "284": "▁c",
+  "285": "▁que",
+  "286": "▁à",
+  "287": "es",
+  "288": "▁un",
+  "289": "▁pas",
+  "290": "▁les",
+  "291": "▁qui",
+  "292": "▁il",
+  "293": "▁vous",
+  "294": "▁des",
+  "295": "▁qu",
+  "296": "▁pour",
+  "297": "▁par",
+  "298": "▁ça",
+  "299": "▁une",
+  "300": "ant",
+  "301": "ais",
+  "302": "ez",
+  "303": "▁dans",
+  "304": "▁C",
+  "305": "tre",
+  "306": "ir",
+  "307": "elle",
+  "308": "eur",
+  "309": "▁sur",
+  "310": "▁con",
+  "311": "▁Et",
+  "312": "ement",
+  "313": "tion",
+  "314": "té",
+  "315": "▁tout",
+  "316": "mp",
+  "317": "ique",
+  "318": "▁plus",
+  "319": "eux",
+  "320": "▁dé",
+  "321": "▁fait",
+  "322": "qu",
+  "323": "▁comme",
+  "324": "ac",
+  "325": "ait",
+  "326": "che",
+  "327": "▁mais",
+  "328": "que",
+  "329": "ul",
+  "330": "▁avec",
+  "331": "▁bien",
+  "332": "age",
+  "333": "▁mon",
+  "334": "end",
+  "335": "▁faire",
+  "336": "▁être",
+  "337": "ver",
+  "338": "▁peu",
+  "339": "▁même",
+  "340": "tra",
+  "341": "cha",
+  "342": "▁nous",
+  "343": "▁donc",
+  "344": "▁sont",
+  "345": "▁moi",
+  "346": "ille",
+  "347": "ff",
+  "348": "▁ex",
+  "349": "▁Il",
+  "350": "▁très",
+  "351": "▁cette",
+  "352": "im",
+  "353": "ité",
+  "354": "▁dire",
+  "355": "▁peut",
+  "356": "ance",
+  "357": "aire",
+  "358": "mé",
+  "359": "▁app",
+  "360": "▁aussi",
+  "361": "▁petit",
+  "362": "aux",
+  "363": "▁parce",
+  "364": "onne",
+  "365": "mb",
+  "366": "man",
+  "367": "▁On",
+  "368": "▁quand",
+  "369": "▁autre",
+  "370": "ô",
+  "371": "▁chose",
+  "372": "▁puis",
+  "373": "▁était",
+  "374": "ndre",
+  "375": "port",
+  "376": "▁vraiment",
+  "377": "ence",
+  "378": "▁Mais",
+  "379": "î",
+  "380": "▁avoir",
+  "381": "form",
+  "382": "▁faut",
+  "383": "ign",
+  "384": "▁où",
+  "385": "près",
+  "386": "▁beaucoup",
+  "387": "ture",
+  "388": "û",
+  "389": "â",
+  "390": "<hu-HU>",
+  "391": "▁az",
+  "392": "ok",
+  "393": "gy",
+  "394": "ek",
+  "395": "és",
+  "396": "em",
+  "397": "▁is",
+  "398": "os",
+  "399": "ban",
+  "400": "ik",
+  "401": "cs",
+  "402": "él",
+  "403": "ér",
+  "404": "nek",
+  "405": "ség",
+  "406": "én",
+  "407": "hat",
+  "408": "▁ha",
+  "409": "ép",
+  "410": "vé",
+  "411": "leg",
+  "412": "▁ami",
+  "413": "É",
+  "414": "rend",
+  "415": "<hr-HR>",
+  "416": "▁bi",
+  "417": "▁sa",
+  "418": "ru",
+  "419": "go",
+  "420": "sti",
+  "421": "▁pri",
+  "422": "ima",
+  "423": "nu",
+  "424": "▁pre",
+  "425": "nja",
+  "426": "zi",
+  "427": "vr",
+  "428": "ca",
+  "429": "ba",
+  "430": "<it-IT>",
+  "431": "▁di",
+  "432": "▁che",
+  "433": "co",
+  "434": "▁per",
+  "435": "▁non",
+  "436": "do",
+  "437": "gli",
+  "438": "so",
+  "439": "amo",
+  "440": "sa",
+  "441": "ndo",
+  "442": "fi",
+  "443": "pi",
+  "444": "nti",
+  "445": "tto",
+  "446": "tro",
+  "447": "▁fa",
+  "448": "chi",
+  "449": "▁qua",
+  "450": "bi",
+  "451": "▁del",
+  "452": "pe",
+  "453": "ssi",
+  "454": "▁sono",
+  "455": "nte",
+  "456": "sso",
+  "457": "nta",
+  "458": "cu",
+  "459": "ggi",
+  "460": "sto",
+  "461": "gra",
+  "462": "spe",
+  "463": "cco",
+  "464": "nde",
+  "465": "fe",
+  "466": "pu",
+  "467": "zza",
+  "468": "sci",
+  "469": "▁poi",
+  "470": "sco",
+  "471": "stra",
+  "472": "▁quel",
+  "473": "qui",
+  "474": "sse",
+  "475": "▁inter",
+  "476": "sce",
+  "477": "▁com",
+  "478": "▁comp",
+  "479": "▁prima",
+  "480": "▁imp",
+  "481": "▁cui",
+  "482": "scri",
+  "483": "<lt-LT>",
+  "484": "ai",
+  "485": "▁ir",
+  "486": "as",
+  "487": "▁tai",
+  "488": "uo",
+  "489": "tin",
+  "490": "▁vis",
+  "491": "ly",
+  "492": "gal",
+  "493": "jo",
+  "494": "tar",
+  "495": "yb",
+  "496": "▁Ir",
+  "497": "▁Tai",
+  "498": "▁nu",
+  "499": "▁mes",
+  "500": "▁pra",
+  "501": "<lv-LV>",
+  "502": "▁pie",
+  "503": "ies",
+  "504": "ot",
+  "505": "▁Un",
+  "506": "▁Ta",
+  "507": "▁lie",
+  "508": "dar",
+  "509": "sim",
+  "510": "dien",
+  "511": "gan",
+  "512": "▁ap",
+  "513": "▁nav",
+  "514": "▁cit",
+  "515": "▁Ne",
+  "516": "zin",
+  "517": "reiz",
+  "518": "▁20",
+  "519": "<nl-NL>",
+  "520": "▁dat",
+  "521": "▁we",
+  "522": "▁En",
+  "523": "▁dan",
+  "524": "▁met",
+  "525": "▁wat",
+  "526": "der",
+  "527": "ui",
+  "528": "den",
+  "529": "op",
+  "530": "oor",
+  "531": "▁of",
+  "532": "ven",
+  "533": "acht",
+  "534": "ond",
+  "535": "▁dit",
+  "536": "▁wil",
+  "537": "▁We",
+  "538": "vol",
+  "539": "llen",
+  "540": "▁Dan",
+  "541": "▁Nou",
+  "542": "<pl-PL>",
+  "543": "nie",
+  "544": "wa",
+  "545": "my",
+  "546": "cie",
+  "547": "nia",
+  "548": "wo",
+  "549": "rze",
+  "550": "by",
+  "551": "za",
+  "552": "dy",
+  "553": "ry",
+  "554": "mie",
+  "555": "pie",
+  "556": "uje",
+  "557": "▁bo",
+  "558": "bie",
+  "559": "sy",
+  "560": "▁Po",
+  "561": "ski",
+  "562": "▁No",
+  "563": "▁Na",
+  "564": "▁musi",
+  "565": "<pt-BR>",
+  "566": "pt",
+  "567": "-",
+  "568": "▁eu",
+  "569": "▁para",
+  "570": "▁pe",
+  "571": "▁tem",
+  "572": "▁O",
+  "573": "pre",
+  "574": "ria",
+  "575": "▁fo",
+  "576": "mos",
+  "577": "bo",
+  "578": "▁por",
+  "579": "▁Eu",
+  "580": "be",
+  "581": "▁esse",
+  "582": "ente",
+  "583": "▁mas",
+  "584": "qua",
+  "585": "fica",
+  "586": "▁vou",
+  "587": "▁Se",
+  "588": "▁fala",
+  "589": "...",
+  "590": "ço",
+  "591": "▁Por",
+  "592": "▁Co",
+  "593": "iza",
+  "594": "<ro-RO>",
+  "595": "▁mai",
+  "596": "sc",
+  "597": "are",
+  "598": "▁din",
+  "599": "rea",
+  "600": "du",
+  "601": "▁M",
+  "602": "▁fac",
+  "603": "lor",
+  "604": "▁mult",
+  "605": "per",
+  "606": "cur",
+  "607": "tor",
+  "608": "inte",
+  "609": "▁sau",
+  "610": "tat",
+  "611": "ori",
+  "612": "▁prim",
+  "613": "▁lui",
+  "614": "▁sub",
+  "615": "▁prin",
+  "616": "▁alt",
+  "617": "stru",
+  "618": "fer",
+  "619": "<ru-RU>",
+  "620": "<sk-SK>",
+  "621": "sk",
+  "622": "▁aj",
+  "623": "ob",
+  "624": "ali",
+  "625": "nov",
+  "626": "rov",
+  "627": "rob",
+  "628": "▁spo",
+  "629": "<sl-SL>",
+  "630": "sl",
+  "631": "udi",
+  "632": "del",
+  "633": "▁sem",
+  "634": "▁samo",
+  "635": "▁let",
+  "636": "▁Pre",
+  "637": "▁prot",
+  "638": "▁internet",
+  "639": "▁film",
+  "640": "<sv-SE>",
+  "641": "▁att",
+  "642": "▁inte",
+  "643": "▁av",
+  "644": "all",
+  "645": "era",
+  "646": "pp",
+  "647": "sam",
+  "648": "▁vill",
+  "649": "erna",
+  "650": "ande",
+  "651": "ade",
+  "652": "▁hur",
+  "653": "bil",
+  "654": "▁min",
+  "655": "lev",
+  "656": "land",
+  "657": "text",
+  "658": "person",
+  "659": "<uk-UA>",
+  "660": "<ar-AR>",
+  "661": "'",
+  "662": "▁H",
+  "663": "▁D",
+  "664": "aus",
+  "665": "▁N",
+  "666": "▁Be",
+  "667": "mm",
+  "668": "ab",
+  "669": "▁Er",
+  "670": "ssen",
+  "671": "hl",
+  "672": "rie",
+  "673": "▁An",
+  "674": "rau",
+  "675": "▁So",
+  "676": "▁Aus",
+  "677": "<en-GB>",
+  "678": "▁can",
+  "679": "ed",
+  "680": "ay",
+  "681": "th",
+  "682": "ic",
+  "683": "hi",
+  "684": "▁not",
+  "685": "ight",
+  "686": "ex",
+  "687": "▁four",
+  "688": "▁great",
+  "689": "▁six",
+  "690": "ill",
+  "691": "▁but",
+  "692": "▁got",
+  "693": "▁don",
+  "694": "▁out",
+  "695": "▁from",
+  "696": "▁fine",
+  "697": "able",
+  "698": "ever",
+  "699": "▁month",
+  "700": "▁question",
+  "701": "▁possible",
+  "702": "▁moment",
+  "703": "▁change",
+  "704": "<en-US>",
+  "705": "ow",
+  "706": "ive",
+  "707": "ate",
+  "708": "ad",
+  "709": "ect",
+  "710": "ng",
+  "711": "ther",
+  "712": "act",
+  "713": "ist",
+  "714": "▁part",
+  "715": "side",
+  "716": "\"",
+  "717": "<es-ES>",
+  "718": "ción",
+  "719": "res",
+  "720": "▁La",
+  "721": "dos",
+  "722": "▁las",
+  "723": "men",
+  "724": "par",
+  "725": "rio",
+  "726": "cho",
+  "727": "ble",
+  "728": "▁Ca",
+  "729": "▁Su",
+  "730": "▁son",
+  "731": "▁Con",
+  "732": "ones",
+  "733": "ecto",
+  "734": "▁Com",
+  "735": "<es-US>",
+  "736": "cia",
+  "737": "▁Y",
+  "738": "ron",
+  "739": "les",
+  "740": "bu",
+  "741": "<fr-CA>",
+  "742": "ré",
+  "743": "ée",
+  "744": "▁Les",
+  "745": "our",
+  "746": "▁Ce",
+  "747": "com",
+  "748": "▁Elle",
+  "749": "▁Cet",
+  "750": "ux",
+  "751": "ale",
+  "752": "ction",
+  "753": "▁pré",
+  "754": "▁deux",
+  "755": "if",
+  "756": "lé",
+  "757": "ère",
+  "758": "ière",
+  "759": "iste",
+  "760": "▁parti",
+  "761": "▁été",
+  "762": "avec",
+  "763": "jour",
+  "764": "app",
+  "765": "cul",
+  "766": "▁égale",
+  "767": "aine",
+  "768": "gue",
+  "769": "▁nombre",
+  "770": "▁étai",
+  "771": "tout",
+  "772": "▁grand",
+  "773": "▁commun",
+  "774": "Une",
+  "775": "œ",
+  "776": "ï",
+  "777": "À",
+  "778": "Â",
+  "779": "av",
+  "780": "ouv",
+  "781": "êt",
+  "782": "ois",
+  "783": "pri",
+  "784": "voir",
+  "785": "sion",
+  "786": "ix",
+  "787": "était",
+  "788": "ard",
+  "789": "aient",
+  "790": "<he-IL>",
+  "791": "<hi-IN>",
+  "792": "!",
+  "793": "<ja-JP>",
+  "794": "<ko-KR>",
+  "795": "<nb-NO>",
+  "796": "▁ble",
+  "797": "opp",
+  "798": "▁Han",
+  "799": "▁han",
+  "800": "<nn-NO>",
+  "801": "eg",
+  "802": "kk",
+  "803": "dig",
+  "804": "ord",
+  "805": "▁tru",
+  "806": "▁sei",
+  "807": "ller",
+  "808": "▁fle",
+  "809": "car",
+  "810": "▁foi",
+  "811": "ito",
+  "812": "ram",
+  "813": "fa",
+  "814": "▁meu",
+  "815": "nça",
+  "816": "▁mil",
+  "817": "▁passa",
+  "818": "▁casa",
+  "819": "<pt-PT>",
+  "820": "das",
+  "821": "▁Pa",
+  "822": "forma",
+  "823": "mar",
+  "824": "▁entre",
+  "825": "fun",
+  "826": "▁grande",
+  "827": "▁Este",
+  "828": "var",
+  "829": "<th-TH>",
+  "830": "<tr-TR>",
+  "831": "ş",
+  "832": "ğ",
+  "833": "ya",
+  "834": "▁ve",
+  "835": "lar",
+  "836": "ler",
+  "837": "ye",
+  "838": "▁bu",
+  "839": "lan",
+  "840": "ara",
+  "841": "▁Bu",
+  "842": "inde",
+  "843": "yo",
+  "844": "<zh-CN>",
+  "845": "<vi-VN>",
+  "846": "▁t",
+  "847": "nh",
+  "848": "<blank>"
+}
\ No newline at end of file
diff --git a/fr/4480ms/decoder.mlmodelc/analytics/coremldata.bin b/fr/4480ms/decoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..65b800e186dcb664ec3d46edb11fa43215dce1f7
--- /dev/null
+++ b/fr/4480ms/decoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d3becf0abd7a7efd30577f33f4144cfcb8c08155fdd62903eb1a45d7b02b87c
+size 243
diff --git a/fr/4480ms/decoder.mlmodelc/coremldata.bin b/fr/4480ms/decoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..b50e1931d2e3334018b12d034c8ff8a57896a265
--- /dev/null
+++ b/fr/4480ms/decoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa812bb65dd2a3bef6acf584b2abd5d0f26f4d09afaccf6c5dfd41e630d0fd1b
+size 433
diff --git a/fr/4480ms/decoder.mlmodelc/model.mil b/fr/4480ms/decoder.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..0de41956c1a2cac40f0629b26d6b60b23d5fe6b3
--- /dev/null
+++ b/fr/4480ms/decoder.mlmodelc/model.mil
@@ -0,0 +1,64 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            int32 y_axis_0 = const()[name = string("y_axis_0"), val = int32(0)];
+            int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)];
+            bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [849, 640]> module_prediction_embed_weight_to_fp16 = const()[name = string("module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_8")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)];
+            int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)];
+            string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")];
+            int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)];
+            int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)];
+            string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            string input_lstm_layer_0_direction_0 = const()[name = string("input_lstm_layer_0_direction_0"), val = string("forward")];
+            bool input_lstm_layer_0_output_sequence_0 = const()[name = string("input_lstm_layer_0_output_sequence_0"), val = bool(true)];
+            string input_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")];
+            string input_lstm_layer_0_cell_activation_0 = const()[name = string("input_lstm_layer_0_cell_activation_0"), val = string("tanh")];
+            string input_lstm_layer_0_activation_0 = const()[name = string("input_lstm_layer_0_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1086848)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4363712)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7640576)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = string("transpose_2")];
+            tensor<fp16, [1, 1, 640]> input_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_lstm_layer_0_cast_fp16_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_lstm_layer_0_cell_activation_0, direction = input_lstm_layer_0_direction_0, initial_c = input_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_lstm_c0_squeeze_cast_fp16")];
+            string input_direction_0 = const()[name = string("input_direction_0"), val = string("forward")];
+            bool input_output_sequence_0 = const()[name = string("input_output_sequence_0"), val = bool(true)];
+            string input_recurrent_activation_0 = const()[name = string("input_recurrent_activation_0"), val = string("sigmoid")];
+            string input_cell_activation_0 = const()[name = string("input_cell_activation_0"), val = string("tanh")];
+            string input_activation_0 = const()[name = string("input_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7645760)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10922624)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14199488)))];
+            tensor<fp16, [1, 1, 640]> input_cast_fp16_0, tensor<fp16, [1, 640]> input_cast_fp16_1, tensor<fp16, [1, 640]> input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_3_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_lstm_layer_0_cast_fp16_0)[name = string("input_cast_fp16")];
+            int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_cast_fp16_1, input_cast_fp16_1))[name = string("obj_3_cast_fp16")];
+            string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_lstm_layer_0_cast_fp16_2, input_cast_fp16_2))[name = string("obj_cast_fp16")];
+            string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
+            string transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = string("transpose_0_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp16, [1, 640, 1]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = string("transpose_1")];
+            tensor<fp32, [1, 640, 1]> decoder_out = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = string("cast_3")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")];
+        } -> (decoder_out, h_out, c_out);
+}
\ No newline at end of file
diff --git a/fr/4480ms/decoder.mlmodelc/weights/weight.bin b/fr/4480ms/decoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..82ea1696ddad052267574c52c3e3b0bccde479f5
--- /dev/null
+++ b/fr/4480ms/decoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ef2aeef85ed664eb60850330a014c5c299325ba1047d690654dc7ffa870e95a
+size 14204672
diff --git a/fr/4480ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/4480ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..c09a4248bce6ae30d0023f4ec349ae462e19aad7
--- /dev/null
+++ b/fr/4480ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c82c0fbdef3af45d82f9b1fe84c0c822154020a5b3cd0e5070d33e9ef0beeca6
+size 10359
diff --git a/fr/4480ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/4480ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..82ea1696ddad052267574c52c3e3b0bccde479f5
--- /dev/null
+++ b/fr/4480ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ef2aeef85ed664eb60850330a014c5c299325ba1047d690654dc7ffa870e95a
+size 14204672
diff --git a/fr/4480ms/decoder.mlpackage/Manifest.json b/fr/4480ms/decoder.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..80bff9abaf12e957f68dbde257eaf82079fed11b
--- /dev/null
+++ b/fr/4480ms/decoder.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "08617906-EB26-428B-BCF5-BC6013603E2C": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "89EF9907-E7AF-4633-B6D7-C8D472B636BC": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "89EF9907-E7AF-4633-B6D7-C8D472B636BC"
+}
diff --git a/fr/4480ms/decoder_joint.mlmodelc/analytics/coremldata.bin b/fr/4480ms/decoder_joint.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..22debc561a233f97bd7fde99dc9d1a31fb0231de
--- /dev/null
+++ b/fr/4480ms/decoder_joint.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f8367de6fb06c7b8053f5ecbbc8e4c7e5100f85f1e1c9ba6e9d9b3e31ecca3d
+size 243
diff --git a/fr/4480ms/decoder_joint.mlmodelc/coremldata.bin b/fr/4480ms/decoder_joint.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..8ccbac6c5ffcbc4faea5c565d772bcf81225e38b
--- /dev/null
+++ b/fr/4480ms/decoder_joint.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f20369afbade05ad03e10dbc687e2f1e1f4f3430bed0a658f03cbd7693296e2b
+size 454
diff --git a/fr/4480ms/decoder_joint.mlmodelc/model.mil b/fr/4480ms/decoder_joint.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..1438ac2ef0ec1f9fc664a6752938e369f7de8e9d
--- /dev/null
+++ b/fr/4480ms/decoder_joint.mlmodelc/model.mil
@@ -0,0 +1,83 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [1, 1024, 1]> encoder, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            int32 y_axis_0 = const()[name = string("y_axis_0"), val = int32(0)];
+            int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)];
+            bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [849, 640]> decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_9")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)];
+            int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)];
+            string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_8")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")];
+            int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)];
+            int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)];
+            string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")];
+            bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)];
+            string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")];
+            string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1086848)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4363712)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7640576)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = string("transpose_4")];
+            tensor<fp16, [1, 1, 640]> input_5_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")];
+            string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")];
+            bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)];
+            string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")];
+            string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7645760)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10922624)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14199488)))];
+            tensor<fp16, [1, 1, 640]> input_5_cast_fp16_0, tensor<fp16, [1, 640]> input_5_cast_fp16_1, tensor<fp16, [1, 640]> input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")];
+            int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")];
+            string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")];
+            string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [3]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [3]> input_7_perm_0 = const()[name = string("input_7_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [640, 1024]> joint_module_enc_weight_to_fp16 = const()[name = string("joint_module_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14204672)))];
+            tensor<fp16, [640]> joint_module_enc_bias_to_fp16 = const()[name = string("joint_module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15515456)))];
+            tensor<fp16, [1, 1024, 1]> encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_4")];
+            tensor<fp16, [1, 1, 1024]> input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = encoder_to_fp16)[name = string("transpose_2")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15516800)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16336064)))];
+            tensor<fp16, [1, 1, 640]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_3")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_79_axes_0 = const()[name = string("op_79_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_79_cast_fp16 = expand_dims(axes = var_79_axes_0, x = linear_0_cast_fp16)[name = string("op_79_cast_fp16")];
+            tensor<int32, [1]> var_80_axes_0 = const()[name = string("op_80_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_80_cast_fp16 = expand_dims(axes = var_80_axes_0, x = linear_1_cast_fp16)[name = string("op_80_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_11_cast_fp16 = add(x = var_79_cast_fp16, y = var_80_cast_fp16)[name = string("input_11_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_13_cast_fp16 = relu(x = input_11_cast_fp16)[name = string("input_13_cast_fp16")];
+            tensor<fp16, [849, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16337408)))];
+            tensor<fp16, [849]> joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [849]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17424192)))];
+            tensor<fp16, [1, 1, 1, 849]> linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_2_cast_fp16")];
+            string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 1, 1, 849]> logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_3")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_5")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_6")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")];
+        } -> (logits, h_out, c_out);
+}
\ No newline at end of file
diff --git a/fr/4480ms/decoder_joint.mlmodelc/weights/weight.bin b/fr/4480ms/decoder_joint.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..7e387cabbe194b0f606c14bdf2b03b19d697a8f3
--- /dev/null
+++ b/fr/4480ms/decoder_joint.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39ce7eecdfb6a432f391c931412d9a5cd15fb0ef85562ef0c207e4a18fe6b50e
+size 17425954
diff --git a/fr/4480ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/4480ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..8de2f672a8b40625e6b0f13654d4cba788880b07
--- /dev/null
+++ b/fr/4480ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72ce6705a73b8b0488f31dd35a385908473aff8bef0ffe290f56bb51068a1d26
+size 13745
diff --git a/fr/4480ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/4480ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..7e387cabbe194b0f606c14bdf2b03b19d697a8f3
--- /dev/null
+++ b/fr/4480ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39ce7eecdfb6a432f391c931412d9a5cd15fb0ef85562ef0c207e4a18fe6b50e
+size 17425954
diff --git a/fr/4480ms/decoder_joint.mlpackage/Manifest.json b/fr/4480ms/decoder_joint.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..967aa622f9a5ac2f95978e9feda32d61c8a73643
--- /dev/null
+++ b/fr/4480ms/decoder_joint.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "8FCEBC80-AF4D-42E1-B70D-96BB47913BC7": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "A77788EC-CA36-4DF3-BB1B-E6182E0ABF28": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "8FCEBC80-AF4D-42E1-B70D-96BB47913BC7"
+}
diff --git a/fr/4480ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin b/fr/4480ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..425b9fbc36cf7602d11ebba8a2e29be972229238
--- /dev/null
+++ b/fr/4480ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:878825967d0c03e7b016222a7b824badb35d1d7e134ad3a973abc00dc5fe353a
+size 243
diff --git a/fr/4480ms/decoder_joint_noencproj.mlmodelc/coremldata.bin b/fr/4480ms/decoder_joint_noencproj.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..542a84cdba5b54648069d8ce056bb885171f50e6
--- /dev/null
+++ b/fr/4480ms/decoder_joint_noencproj.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06ddd4ebfa5cbca8400ba528fb93ce16a49ccd898afa2b536795b0aafa77af61
+size 519
diff --git a/fr/4480ms/decoder_joint_noencproj.mlmodelc/model.mil b/fr/4480ms/decoder_joint_noencproj.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..ef9fa8f205f75bf414b61d0b622d2976b441fe5b
--- /dev/null
+++ b/fr/4480ms/decoder_joint_noencproj.mlmodelc/model.mil
@@ -0,0 +1,91 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios18>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [1, 1, 640]> encoder_proj, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)];
+            bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [849, 640]> decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")];
+            string cast_1_dtype_0 = const()[name = string("cast_1_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_10")];
+            tensor<int32, [1, 1]> cast_1 = cast(dtype = cast_1_dtype_0, x = token_to_int16)[name = string("cast_9")];
+            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
+            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(849)];
+            tensor<int32, [1, 1]> add_2 = add(x = cast_1, y = slice_by_index_0)[name = string("add_2")];
+            tensor<int32, [1, 1]> select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = string("select_0")];
+            int32 y_cast_fp16_cast_uint16_axis_0 = const()[name = string("y_cast_fp16_cast_uint16_axis_0"), val = int32(0)];
+            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
+            tensor<int16, [1, 1]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_8")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)];
+            int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)];
+            string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")];
+            int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)];
+            int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)];
+            string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")];
+            bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)];
+            string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")];
+            string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1086848)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4363712)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7640576)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = string("transpose_3")];
+            tensor<fp16, [1, 1, 640]> input_5_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")];
+            string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")];
+            bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)];
+            string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")];
+            string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7645760)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10922624)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14199488)))];
+            tensor<fp16, [1, 1, 640]> input_5_cast_fp16_0, tensor<fp16, [1, 640]> input_5_cast_fp16_1, tensor<fp16, [1, 640]> input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")];
+            int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")];
+            string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")];
+            string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [3]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14204672)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15023936)))];
+            tensor<fp16, [1, 1, 640]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_2")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [1]> f_axes_0 = const()[name = string("f_axes_0"), val = tensor<int32, [1]>([2])];
+            string encoder_proj_to_fp16_dtype_0 = const()[name = string("encoder_proj_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 1, 640]> encoder_proj_to_fp16 = cast(dtype = encoder_proj_to_fp16_dtype_0, x = encoder_proj)[name = string("cast_3")];
+            tensor<fp16, [1, 1, 1, 640]> f_cast_fp16 = expand_dims(axes = f_axes_0, x = encoder_proj_to_fp16)[name = string("f_cast_fp16")];
+            tensor<int32, [1]> g_axes_0 = const()[name = string("g_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> g_cast_fp16 = expand_dims(axes = g_axes_0, x = linear_0_cast_fp16)[name = string("g_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_9_cast_fp16 = add(x = f_cast_fp16, y = g_cast_fp16)[name = string("input_9_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_11_cast_fp16 = relu(x = input_9_cast_fp16)[name = string("input_11_cast_fp16")];
+            tensor<fp16, [849, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15025280)))];
+            tensor<fp16, [849]> joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [849]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16112064)))];
+            tensor<fp16, [1, 1, 1, 849]> linear_1_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_1_cast_fp16")];
+            int32 var_83 = const()[name = string("op_83"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1, 849]> var_85_softmax_cast_fp16 = softmax(axis = var_83, x = linear_1_cast_fp16)[name = string("op_85_softmax_cast_fp16")];
+            fp32 var_85_epsilon_0 = const()[name = string("op_85_epsilon_0"), val = fp32(0x1p-149)];
+            tensor<fp16, [1, 1, 1, 849]> var_85_cast_fp16 = log(epsilon = var_85_epsilon_0, x = var_85_softmax_cast_fp16)[name = string("op_85_cast_fp16")];
+            string var_85_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_85_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 1, 1, 849]> logits = cast(dtype = var_85_cast_fp16_to_fp32_dtype_0, x = var_85_cast_fp16)[name = string("cast_2")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")];
+        } -> (logits, h_out, c_out);
+}
\ No newline at end of file
diff --git a/fr/4480ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin b/fr/4480ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..e37358fc0d446e079fe87526671dc411c4b37591
--- /dev/null
+++ b/fr/4480ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c691d3004f2ac6a956208569a29d899a7f8e5030087d53fbd488785fa0e8efad
+size 16113826
diff --git a/fr/4480ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/4480ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..b701505b2faf5efd4a3c1024077f4834b1f94d37
--- /dev/null
+++ b/fr/4480ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4196a821e716db75328fdd2f99f83214ecd9e6e6481a96a88907fe1c4232f143
+size 14630
diff --git a/fr/4480ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/4480ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..e37358fc0d446e079fe87526671dc411c4b37591
--- /dev/null
+++ b/fr/4480ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c691d3004f2ac6a956208569a29d899a7f8e5030087d53fbd488785fa0e8efad
+size 16113826
diff --git a/fr/4480ms/decoder_joint_noencproj.mlpackage/Manifest.json b/fr/4480ms/decoder_joint_noencproj.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..4709121d3572643a899871aa076a46f4c02a7ef8
--- /dev/null
+++ b/fr/4480ms/decoder_joint_noencproj.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "1D4938C1-F035-4D52-9376-C0FB72335508": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "BF28BA0A-268B-4C46-AB97-C0173D8B55F0": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "BF28BA0A-268B-4C46-AB97-C0173D8B55F0"
+}
diff --git a/fr/4480ms/encoder.mlmodelc/analytics/coremldata.bin b/fr/4480ms/encoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..0ad96ceb53ea0de319036385ee673f9d5c4c247e
--- /dev/null
+++ b/fr/4480ms/encoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e3bb93b7dca0a3022e499530cc2b070982c8a0b0afffed7b0d9c9f6cf943e091
+size 243
diff --git a/fr/4480ms/encoder.mlmodelc/coremldata.bin b/fr/4480ms/encoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..e34d98d530cead4447b2257e18ae93c2e31ec8cc
--- /dev/null
+++ b/fr/4480ms/encoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:651e75f9c559382017203a2ddc3e1c7b3eda97d8f7fdfba6a94c95a44a158700
+size 573
diff --git a/fr/4480ms/encoder.mlmodelc/model.mil b/fr/4480ms/encoder.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..ac7ff9f2bb92afa45e5a2fabeefa184a92678a46
--- /dev/null
+++ b/fr/4480ms/encoder.mlmodelc/model.mil
@@ -0,0 +1,4321 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})]
+{
+    func main<ios18>(tensor<fp32, [1, 24, 42, 1024]> cache_channel, tensor<int32, [1]> cache_len, tensor<fp32, [1, 24, 1024, 8]> cache_time, tensor<fp32, [1, 128, 457]> mel, tensor<int32, [1]> mel_length, tensor<int32, [1]> prompt_id) {
+            tensor<int32, [4]> value_3_perm_0 = const()[name = string("value_3_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string cache_channel_to_fp16_dtype_0 = const()[name = string("cache_channel_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<int32, [4]> value_5_perm_0 = const()[name = string("value_5_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string cache_time_to_fp16_dtype_0 = const()[name = string("cache_time_to_fp16_dtype_0"), val = string("fp16")];
+            int32 var_58 = const()[name = string("op_58"), val = int32(-1)];
+            int32 var_67 = const()[name = string("op_67"), val = int32(1)];
+            tensor<int32, [3]> x_1_perm_0 = const()[name = string("x_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string mel_to_fp16_dtype_0 = const()[name = string("mel_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<int32, [1]> tensor_1_axes_0 = const()[name = string("tensor_1_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 128, 457]> mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = string("cast_18")];
+            tensor<fp16, [1, 457, 128]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = mel_to_fp16)[name = string("transpose_367")];
+            tensor<fp16, [1, 1, 457, 128]> tensor_1_cast_fp16 = expand_dims(axes = tensor_1_axes_0, x = x_1_cast_fp16)[name = string("tensor_1_cast_fp16")];
+            tensor<int32, [1, 457]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1, 457]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<int32, [1]> var_136_axes_0 = const()[name = string("op_136_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_136 = expand_dims(axes = var_136_axes_0, x = mel_length)[name = string("op_136")];
+            tensor<bool, [1, 457]> time_mask_1 = less(x = expand_dims_0, y = var_136)[name = string("time_mask_1")];
+            tensor<int32, [1]> var_138_axes_0 = const()[name = string("op_138_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 457, 1]> var_138 = expand_dims(axes = var_138_axes_0, x = time_mask_1)[name = string("op_138")];
+            tensor<int32, [3]> var_140_reps_0 = const()[name = string("op_140_reps_0"), val = tensor<int32, [3]>([1, 1, 128])];
+            tensor<bool, [1, 457, 128]> var_140 = tile(reps = var_140_reps_0, x = var_138)[name = string("op_140")];
+            tensor<int32, [1]> var_146_axes_0 = const()[name = string("op_146_axes_0"), val = tensor<int32, [1]>([1])];
+            string cast_4_to_fp16_dtype_0 = const()[name = string("cast_4_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 457, 128]> var_140_to_fp16 = cast(dtype = cast_4_to_fp16_dtype_0, x = var_140)[name = string("cast_17")];
+            tensor<fp16, [1, 1, 457, 128]> var_146_cast_fp16 = expand_dims(axes = var_146_axes_0, x = var_140_to_fp16)[name = string("op_146_cast_fp16")];
+            tensor<fp16, [1, 1, 457, 128]> input_1_cast_fp16 = mul(x = tensor_1_cast_fp16, y = var_146_cast_fp16)[name = string("input_1_cast_fp16")];
+            tensor<int32, [8]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("constant")];
+            fp16 const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 460, 131]> input_3_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")];
+            string tensor_3_pad_type_0 = const()[name = string("tensor_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_3_strides_0 = const()[name = string("tensor_3_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, [4]> tensor_3_pad_0 = const()[name = string("tensor_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_3_dilations_0 = const()[name = string("tensor_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 tensor_3_groups_0 = const()[name = string("tensor_3_groups_0"), val = int32(1)];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_0_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1984))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4352))))[name = string("encoder_pre_encode_conv_0_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4928)))];
+            tensor<fp16, [1, 256, 229, 65]> tensor_3_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = tensor_3_dilations_0, groups = tensor_3_groups_0, pad = tensor_3_pad_0, pad_type = tensor_3_pad_type_0, strides = tensor_3_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("tensor_3_cast_fp16")];
+            string cast_2_to_fp16_dtype_0 = const()[name = string("cast_2_to_fp16_dtype_0"), val = string("fp16")];
+            fp16 var_159_promoted_to_fp16 = const()[name = string("op_159_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> mel_length_to_fp16 = cast(dtype = cast_2_to_fp16_dtype_0, x = mel_length)[name = string("cast_16")];
+            tensor<fp16, [1]> var_160_cast_fp16 = add(x = mel_length_to_fp16, y = var_159_promoted_to_fp16)[name = string("op_160_cast_fp16")];
+            fp16 var_161_promoted_to_fp16 = const()[name = string("op_161_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> var_162_cast_fp16 = add(x = var_160_cast_fp16, y = var_161_promoted_to_fp16)[name = string("op_162_cast_fp16")];
+            fp16 var_163_promoted_to_fp16 = const()[name = string("op_163_promoted_to_fp16"), val = fp16(0x1.8p+1)];
+            tensor<fp16, [1]> var_164_cast_fp16 = sub(x = var_162_cast_fp16, y = var_163_promoted_to_fp16)[name = string("op_164_cast_fp16")];
+            fp16 var_55_promoted_to_fp16 = const()[name = string("op_55_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> floor_div_0_cast_fp16 = floor_div(x = var_164_cast_fp16, y = var_55_promoted_to_fp16)[name = string("floor_div_0_cast_fp16")];
+            fp16 var_166_promoted_to_fp16 = const()[name = string("op_166_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_3_cast_fp16 = add(x = floor_div_0_cast_fp16, y = var_166_promoted_to_fp16)[name = string("current_lengths_3_cast_fp16")];
+            string cast_5_dtype_0 = const()[name = string("cast_5_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 229]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1, 229]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5504)))];
+            tensor<int32, [1]> var_175_axes_0 = const()[name = string("op_175_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_3_cast_fp16_to_int32 = cast(dtype = cast_5_dtype_0, x = current_lengths_3_cast_fp16)[name = string("cast_15")];
+            tensor<int32, [1, 1]> var_175 = expand_dims(axes = var_175_axes_0, x = current_lengths_3_cast_fp16_to_int32)[name = string("op_175")];
+            tensor<bool, [1, 229]> time_mask_3 = less(x = expand_dims_1, y = var_175)[name = string("time_mask_3")];
+            tensor<int32, [1]> var_177_axes_0 = const()[name = string("op_177_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 229, 1]> var_177 = expand_dims(axes = var_177_axes_0, x = time_mask_3)[name = string("op_177")];
+            tensor<int32, [3]> var_179_reps_0 = const()[name = string("op_179_reps_0"), val = tensor<int32, [3]>([1, 1, 65])];
+            tensor<bool, [1, 229, 65]> var_179 = tile(reps = var_179_reps_0, x = var_177)[name = string("op_179")];
+            tensor<int32, [1]> var_185_axes_0 = const()[name = string("op_185_axes_0"), val = tensor<int32, [1]>([1])];
+            string cast_6_to_fp16_dtype_0 = const()[name = string("cast_6_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 229, 65]> var_179_to_fp16 = cast(dtype = cast_6_to_fp16_dtype_0, x = var_179)[name = string("cast_14")];
+            tensor<fp16, [1, 1, 229, 65]> var_185_cast_fp16 = expand_dims(axes = var_185_axes_0, x = var_179_to_fp16)[name = string("op_185_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_3_reps_0 = const()[name = string("expanded_mask_3_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 229, 65]> expanded_mask_3_cast_fp16 = tile(reps = expanded_mask_3_reps_0, x = var_185_cast_fp16)[name = string("expanded_mask_3_cast_fp16")];
+            tensor<fp16, [1, 256, 229, 65]> input_5_cast_fp16 = mul(x = tensor_3_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<fp16, [1, 256, 229, 65]> tensor_5_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("tensor_5_cast_fp16")];
+            tensor<fp16, [1, 256, 229, 65]> input_7_cast_fp16 = mul(x = tensor_5_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_7_cast_fp16")];
+            tensor<int32, [8]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            string input_9_mode_0 = const()[name = string("input_9_mode_0"), val = string("constant")];
+            fp16 const_23_to_fp16 = const()[name = string("const_23_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 256, 232, 68]> input_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")];
+            string tensor_7_pad_type_0 = const()[name = string("tensor_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_7_strides_0 = const()[name = string("tensor_7_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            int32 tensor_7_groups_0 = const()[name = string("tensor_7_groups_0"), val = int32(256)];
+            tensor<int32, [4]> tensor_7_pad_0 = const()[name = string("tensor_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_7_dilations_0 = const()[name = string("tensor_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6528))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8896))))[name = string("encoder_pre_encode_conv_2_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9472)))];
+            tensor<fp16, [1, 256, 115, 33]> tensor_7_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = tensor_7_dilations_0, groups = tensor_7_groups_0, pad = tensor_7_pad_0, pad_type = tensor_7_pad_type_0, strides = tensor_7_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("tensor_7_cast_fp16")];
+            fp16 var_207_promoted_to_fp16 = const()[name = string("op_207_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> var_208_cast_fp16 = add(x = current_lengths_3_cast_fp16, y = var_207_promoted_to_fp16)[name = string("op_208_cast_fp16")];
+            fp16 var_209_promoted_to_fp16 = const()[name = string("op_209_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> var_210_cast_fp16 = add(x = var_208_cast_fp16, y = var_209_promoted_to_fp16)[name = string("op_210_cast_fp16")];
+            fp16 var_211_promoted_to_fp16 = const()[name = string("op_211_promoted_to_fp16"), val = fp16(0x1.8p+1)];
+            tensor<fp16, [1]> var_212_cast_fp16 = sub(x = var_210_cast_fp16, y = var_211_promoted_to_fp16)[name = string("op_212_cast_fp16")];
+            fp16 var_55_promoted_1_to_fp16 = const()[name = string("op_55_promoted_1_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> floor_div_1_cast_fp16 = floor_div(x = var_212_cast_fp16, y = var_55_promoted_1_to_fp16)[name = string("floor_div_1_cast_fp16")];
+            fp16 var_214_promoted_to_fp16 = const()[name = string("op_214_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_5_cast_fp16 = add(x = floor_div_1_cast_fp16, y = var_214_promoted_to_fp16)[name = string("current_lengths_5_cast_fp16")];
+            string cast_7_dtype_0 = const()[name = string("cast_7_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 115]> expand_dims_2 = const()[name = string("expand_dims_2"), val = tensor<int32, [1, 115]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10048)))];
+            tensor<int32, [1]> var_223_axes_0 = const()[name = string("op_223_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_5_cast_fp16_to_int32 = cast(dtype = cast_7_dtype_0, x = current_lengths_5_cast_fp16)[name = string("cast_13")];
+            tensor<int32, [1, 1]> var_223 = expand_dims(axes = var_223_axes_0, x = current_lengths_5_cast_fp16_to_int32)[name = string("op_223")];
+            tensor<bool, [1, 115]> time_mask_5 = less(x = expand_dims_2, y = var_223)[name = string("time_mask_5")];
+            tensor<int32, [1]> var_225_axes_0 = const()[name = string("op_225_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 115, 1]> var_225 = expand_dims(axes = var_225_axes_0, x = time_mask_5)[name = string("op_225")];
+            tensor<int32, [3]> var_227_reps_0 = const()[name = string("op_227_reps_0"), val = tensor<int32, [3]>([1, 1, 33])];
+            tensor<bool, [1, 115, 33]> var_227 = tile(reps = var_227_reps_0, x = var_225)[name = string("op_227")];
+            tensor<int32, [1]> var_233_axes_0 = const()[name = string("op_233_axes_0"), val = tensor<int32, [1]>([1])];
+            string cast_8_to_fp16_dtype_0 = const()[name = string("cast_8_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 115, 33]> var_227_to_fp16 = cast(dtype = cast_8_to_fp16_dtype_0, x = var_227)[name = string("cast_12")];
+            tensor<fp16, [1, 1, 115, 33]> var_233_cast_fp16 = expand_dims(axes = var_233_axes_0, x = var_227_to_fp16)[name = string("op_233_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_7_reps_0 = const()[name = string("expanded_mask_7_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 115, 33]> expanded_mask_7_cast_fp16 = tile(reps = expanded_mask_7_reps_0, x = var_233_cast_fp16)[name = string("expanded_mask_7_cast_fp16")];
+            tensor<fp16, [1, 256, 115, 33]> input_11_cast_fp16 = mul(x = tensor_7_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_11_cast_fp16")];
+            string tensor_9_pad_type_0 = const()[name = string("tensor_9_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_9_strides_0 = const()[name = string("tensor_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> tensor_9_pad_0 = const()[name = string("tensor_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_9_dilations_0 = const()[name = string("tensor_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 tensor_9_groups_0 = const()[name = string("tensor_9_groups_0"), val = int32(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_3_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10624))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76224))))[name = string("encoder_pre_encode_conv_3_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76800)))];
+            tensor<fp16, [1, 256, 115, 33]> tensor_9_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = tensor_9_dilations_0, groups = tensor_9_groups_0, pad = tensor_9_pad_0, pad_type = tensor_9_pad_type_0, strides = tensor_9_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("tensor_9_cast_fp16")];
+            tensor<fp16, [1, 256, 115, 33]> input_13_cast_fp16 = mul(x = tensor_9_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_13_cast_fp16")];
+            tensor<fp16, [1, 256, 115, 33]> tensor_11_cast_fp16 = relu(x = input_13_cast_fp16)[name = string("tensor_11_cast_fp16")];
+            tensor<fp16, [1, 256, 115, 33]> input_15_cast_fp16 = mul(x = tensor_11_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_15_cast_fp16")];
+            tensor<int32, [8]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            string input_17_mode_0 = const()[name = string("input_17_mode_0"), val = string("constant")];
+            fp16 const_41_to_fp16 = const()[name = string("const_41_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 256, 118, 36]> input_17_cast_fp16 = pad(constant_val = const_41_to_fp16, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")];
+            string tensor_13_pad_type_0 = const()[name = string("tensor_13_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_13_strides_0 = const()[name = string("tensor_13_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            int32 tensor_13_groups_0 = const()[name = string("tensor_13_groups_0"), val = int32(256)];
+            tensor<int32, [4]> tensor_13_pad_0 = const()[name = string("tensor_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_13_dilations_0 = const()[name = string("tensor_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_5_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77376))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79744))))[name = string("encoder_pre_encode_conv_5_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80320)))];
+            tensor<fp16, [1, 256, 58, 17]> tensor_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = tensor_13_dilations_0, groups = tensor_13_groups_0, pad = tensor_13_pad_0, pad_type = tensor_13_pad_type_0, strides = tensor_13_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("tensor_13_cast_fp16")];
+            fp16 var_270_promoted_to_fp16 = const()[name = string("op_270_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> var_271_cast_fp16 = add(x = current_lengths_5_cast_fp16, y = var_270_promoted_to_fp16)[name = string("op_271_cast_fp16")];
+            fp16 var_272_promoted_to_fp16 = const()[name = string("op_272_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> var_273_cast_fp16 = add(x = var_271_cast_fp16, y = var_272_promoted_to_fp16)[name = string("op_273_cast_fp16")];
+            fp16 var_274_promoted_to_fp16 = const()[name = string("op_274_promoted_to_fp16"), val = fp16(0x1.8p+1)];
+            tensor<fp16, [1]> var_275_cast_fp16 = sub(x = var_273_cast_fp16, y = var_274_promoted_to_fp16)[name = string("op_275_cast_fp16")];
+            fp16 var_55_promoted_2_to_fp16 = const()[name = string("op_55_promoted_2_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> floor_div_2_cast_fp16 = floor_div(x = var_275_cast_fp16, y = var_55_promoted_2_to_fp16)[name = string("floor_div_2_cast_fp16")];
+            fp16 var_277_promoted_to_fp16 = const()[name = string("op_277_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_cast_fp16 = add(x = floor_div_2_cast_fp16, y = var_277_promoted_to_fp16)[name = string("current_lengths_cast_fp16")];
+            string cast_9_dtype_0 = const()[name = string("cast_9_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 58]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1, 58]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80896)))];
+            tensor<int32, [1]> var_286_axes_0 = const()[name = string("op_286_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_cast_fp16_to_int32 = cast(dtype = cast_9_dtype_0, x = current_lengths_cast_fp16)[name = string("cast_11")];
+            tensor<int32, [1, 1]> var_286 = expand_dims(axes = var_286_axes_0, x = current_lengths_cast_fp16_to_int32)[name = string("op_286")];
+            tensor<bool, [1, 58]> time_mask = less(x = expand_dims_3, y = var_286)[name = string("time_mask")];
+            tensor<int32, [1]> var_288_axes_0 = const()[name = string("op_288_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 58, 1]> var_288 = expand_dims(axes = var_288_axes_0, x = time_mask)[name = string("op_288")];
+            tensor<int32, [3]> var_290_reps_0 = const()[name = string("op_290_reps_0"), val = tensor<int32, [3]>([1, 1, 17])];
+            tensor<bool, [1, 58, 17]> var_290 = tile(reps = var_290_reps_0, x = var_288)[name = string("op_290")];
+            tensor<int32, [1]> var_296_axes_0 = const()[name = string("op_296_axes_0"), val = tensor<int32, [1]>([1])];
+            string cast_10_to_fp16_dtype_0 = const()[name = string("cast_10_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 58, 17]> var_290_to_fp16 = cast(dtype = cast_10_to_fp16_dtype_0, x = var_290)[name = string("cast_10")];
+            tensor<fp16, [1, 1, 58, 17]> var_296_cast_fp16 = expand_dims(axes = var_296_axes_0, x = var_290_to_fp16)[name = string("op_296_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_13_reps_0 = const()[name = string("expanded_mask_13_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 58, 17]> expanded_mask_13_cast_fp16 = tile(reps = expanded_mask_13_reps_0, x = var_296_cast_fp16)[name = string("expanded_mask_13_cast_fp16")];
+            tensor<fp16, [1, 256, 58, 17]> input_19_cast_fp16 = mul(x = tensor_13_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_19_cast_fp16")];
+            string tensor_15_pad_type_0 = const()[name = string("tensor_15_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_15_strides_0 = const()[name = string("tensor_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> tensor_15_pad_0 = const()[name = string("tensor_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_15_dilations_0 = const()[name = string("tensor_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 tensor_15_groups_0 = const()[name = string("tensor_15_groups_0"), val = int32(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_6_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81216))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146816))))[name = string("encoder_pre_encode_conv_6_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147392)))];
+            tensor<fp16, [1, 256, 58, 17]> tensor_15_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = tensor_15_dilations_0, groups = tensor_15_groups_0, pad = tensor_15_pad_0, pad_type = tensor_15_pad_type_0, strides = tensor_15_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("tensor_15_cast_fp16")];
+            tensor<fp16, [1, 256, 58, 17]> input_21_cast_fp16 = mul(x = tensor_15_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_21_cast_fp16")];
+            tensor<fp16, [1, 256, 58, 17]> tensor_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("tensor_cast_fp16")];
+            tensor<fp16, [1, 256, 58, 17]> x_3_cast_fp16 = mul(x = tensor_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("x_3_cast_fp16")];
+            tensor<int32, [4]> var_330_perm_0 = const()[name = string("op_330_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_331 = const()[name = string("op_331"), val = tensor<int32, [3]>([1, 58, -1])];
+            tensor<fp16, [1, 58, 256, 17]> var_330_cast_fp16 = transpose(perm = var_330_perm_0, x = x_3_cast_fp16)[name = string("transpose_366")];
+            tensor<fp16, [1, 58, 4352]> input_23_cast_fp16 = reshape(shape = var_331, x = var_330_cast_fp16)[name = string("input_23_cast_fp16")];
+            tensor<fp16, [1024, 4352]> encoder_pre_encode_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4352]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147968))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4604480))))[name = string("encoder_pre_encode_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_pre_encode_out_bias_to_fp16 = const()[name = string("encoder_pre_encode_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606592)))];
+            tensor<fp16, [1, 58, 1024]> linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [3]> var_341_begin_0 = const()[name = string("op_341_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_341_end_0 = const()[name = string("op_341_end_0"), val = tensor<int32, [3]>([1, 58, 1024])];
+            tensor<bool, [3]> var_341_end_mask_0 = const()[name = string("op_341_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 56, 1024]> var_341_cast_fp16 = slice_by_index(begin = var_341_begin_0, end = var_341_end_0, end_mask = var_341_end_mask_0, x = linear_0_cast_fp16)[name = string("op_341_cast_fp16")];
+            int32 var_343 = const()[name = string("op_343"), val = int32(2)];
+            tensor<int32, [1]> var_344 = sub(x = current_lengths_cast_fp16_to_int32, y = var_343)[name = string("op_344")];
+            string var_344_promoted_to_fp16_dtype_0 = const()[name = string("op_344_promoted_to_fp16_dtype_0"), val = string("fp16")];
+            fp16 var_61_promoted_to_fp16 = const()[name = string("op_61_promoted_to_fp16"), val = fp16(0x0p+0)];
+            fp16 const_61_to_fp16 = const()[name = string("const_61_to_fp16"), val = fp16(inf)];
+            tensor<fp16, [1]> var_344_to_fp16 = cast(dtype = var_344_promoted_to_fp16_dtype_0, x = var_344)[name = string("cast_9")];
+            tensor<fp16, [1]> clip_0_cast_fp16 = clip(alpha = var_61_promoted_to_fp16, beta = const_61_to_fp16, x = var_344_to_fp16)[name = string("clip_0_cast_fp16")];
+            tensor<int32, [1]> max_audio_length_1 = const()[name = string("max_audio_length_1"), val = tensor<int32, [1]>([56])];
+            fp16 var_360_promoted_to_fp16 = const()[name = string("op_360_promoted_to_fp16"), val = fp16(0x1.5p+5)];
+            tensor<fp16, [1]> padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_360_promoted_to_fp16)[name = string("padding_length_cast_fp16")];
+            int32 const_63 = const()[name = string("const_63"), val = int32(-1)];
+            tensor<int32, [1]> var_362 = mul(x = cache_len, y = const_63)[name = string("op_362")];
+            int32 var_363 = const()[name = string("op_363"), val = int32(42)];
+            tensor<int32, [1]> offset = add(x = var_362, y = var_363)[name = string("offset")];
+            tensor<int32, [1]> var_403_axes_0 = const()[name = string("op_403_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 1]> var_403_cast_fp16 = expand_dims(axes = var_403_axes_0, x = padding_length_cast_fp16)[name = string("op_403_cast_fp16")];
+            tensor<fp16, [1, 98]> var_402_promoted_to_fp16 = const()[name = string("op_402_promoted_to_fp16"), val = tensor<fp16, [1, 98]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4608704)))];
+            tensor<bool, [1, 98]> pad_mask_1_cast_fp16 = less(x = var_402_promoted_to_fp16, y = var_403_cast_fp16)[name = string("pad_mask_1_cast_fp16")];
+            tensor<int32, [1, 98]> expand_dims_5 = const()[name = string("expand_dims_5"), val = tensor<int32, [1, 98]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4609024)))];
+            tensor<int32, [1]> var_409_axes_0 = const()[name = string("op_409_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 1]> var_409 = expand_dims(axes = var_409_axes_0, x = offset)[name = string("op_409")];
+            tensor<bool, [1, 98]> pad_mask_off = greater_equal(x = expand_dims_5, y = var_409)[name = string("pad_mask_off")];
+            tensor<bool, [1, 98]> pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = string("pad_mask_3")];
+            tensor<int32, [1]> var_412_axes_0 = const()[name = string("op_412_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 98]> var_412 = expand_dims(axes = var_412_axes_0, x = pad_mask_3)[name = string("op_412")];
+            tensor<int32, [3]> var_413 = const()[name = string("op_413"), val = tensor<int32, [3]>([1, 98, 1])];
+            tensor<bool, [1, 98, 98]> pad_mask_for_att_mask_1 = tile(reps = var_413, x = var_412)[name = string("pad_mask_for_att_mask_1")];
+            tensor<int32, [3]> var_415_perm_0 = const()[name = string("op_415_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<bool, [1, 98, 98]> var_415 = transpose(perm = var_415_perm_0, x = pad_mask_for_att_mask_1)[name = string("transpose_365")];
+            tensor<bool, [1, 98, 98]> pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_415)[name = string("pad_mask_for_att_mask")];
+            tensor<bool, [1, 98, 98]> const_71 = const()[name = string("const_71"), val = tensor<bool, [1, 98, 98]>([[[true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])];
+            tensor<bool, [1, 98, 98]> att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_71)[name = string("att_mask_9")];
+            tensor<bool, [1, 98, 98]> att_mask = logical_not(x = att_mask_9)[name = string("att_mask")];
+            tensor<bool, [1, 98]> pad_mask_5 = logical_not(x = pad_mask_3)[name = string("pad_mask_5")];
+            tensor<int32, [2]> pad_mask_begin_0 = const()[name = string("pad_mask_begin_0"), val = tensor<int32, [2]>([0, 42])];
+            tensor<int32, [2]> pad_mask_end_0 = const()[name = string("pad_mask_end_0"), val = tensor<int32, [2]>([1, 98])];
+            tensor<bool, [2]> pad_mask_end_mask_0 = const()[name = string("pad_mask_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<bool, [1, 56]> pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = string("pad_mask")];
+            tensor<int32, [3]> mask_9_begin_0 = const()[name = string("mask_9_begin_0"), val = tensor<int32, [3]>([0, 42, 0])];
+            tensor<int32, [3]> mask_9_end_0 = const()[name = string("mask_9_end_0"), val = tensor<int32, [3]>([1, 98, 98])];
+            tensor<bool, [3]> mask_9_end_mask_0 = const()[name = string("mask_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<bool, [1, 56, 98]> mask_9 = slice_by_index(begin = mask_9_begin_0, end = mask_9_end_0, end_mask = mask_9_end_mask_0, x = att_mask)[name = string("mask_9")];
+            tensor<int32, [4]> cache_1_begin_0 = const()[name = string("cache_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_1_end_0 = const()[name = string("cache_1_end_0"), val = tensor<int32, [4]>([1, 1, 42, 1024])];
+            tensor<bool, [4]> cache_1_end_mask_0 = const()[name = string("cache_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_1_squeeze_mask_0 = const()[name = string("cache_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 24, 42, 1024]> cache_channel_to_fp16 = cast(dtype = cache_channel_to_fp16_dtype_0, x = cache_channel)[name = string("cast_8")];
+            tensor<fp16, [24, 1, 42, 1024]> value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = cache_channel_to_fp16)[name = string("transpose_364")];
+            tensor<fp16, [1, 42, 1024]> cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_1_cast_fp16")];
+            tensor<int32, [4]> cache_3_begin_0 = const()[name = string("cache_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_3_end_0 = const()[name = string("cache_3_end_0"), val = tensor<int32, [4]>([1, 1, 1024, 8])];
+            tensor<bool, [4]> cache_3_end_mask_0 = const()[name = string("cache_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_3_squeeze_mask_0 = const()[name = string("cache_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 24, 1024, 8]> cache_time_to_fp16 = cast(dtype = cache_time_to_fp16_dtype_0, x = cache_time)[name = string("cast_7")];
+            tensor<fp16, [24, 1, 1024, 8]> value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = cache_time_to_fp16)[name = string("transpose_363")];
+            tensor<fp16, [1, 1024, 8]> cache_3_cast_fp16 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_3_cast_fp16")];
+            tensor<int32, [1]> input_27_axes_0 = const()[name = string("input_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4609536)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4611648)))];
+            fp16 var_41_to_fp16 = const()[name = string("op_41_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 56, 1024]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_341_cast_fp16)[name = string("input_27_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4613760))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8808128))))[name = string("encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_0_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8816384)))];
+            tensor<fp16, [1, 56, 4096]> linear_1_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = string("input_31_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8824640))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13019008))))[name = string("encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13021120)))];
+            tensor<fp16, [1, 56, 1024]> linear_2_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_2_cast_fp16")];
+            fp16 var_454_to_fp16 = const()[name = string("op_454_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_455_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_454_to_fp16)[name = string("op_455_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_37_cast_fp16 = add(x = var_341_cast_fp16, y = var_455_cast_fp16)[name = string("input_37_cast_fp16")];
+            tensor<int32, [1]> key_1_axes_0 = const()[name = string("key_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13023232)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13025344)))];
+            tensor<fp16, [1, 56, 1024]> key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = string("key_1_cast_fp16")];
+            bool input_39_interleave_0 = const()[name = string("input_39_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_39_cast_fp16 = concat(axis = var_67, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = string("input_39_cast_fp16")];
+            bool var_483_interleave_0 = const()[name = string("op_483_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_483_cast_fp16 = concat(axis = var_67, interleave = var_483_interleave_0, values = key_1_cast_fp16)[name = string("op_483_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13027456))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14076096))))[name = string("encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14078208)))];
+            tensor<fp16, [1, 56, 1024]> linear_3_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized, x = key_1_cast_fp16)[name = string("linear_3_cast_fp16")];
+            tensor<int32, [4]> var_488 = const()[name = string("op_488"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_1_cast_fp16 = reshape(shape = var_488, x = linear_3_cast_fp16)[name = string("q_1_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14080320))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15128960))))[name = string("encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15131072)))];
+            tensor<fp16, [1, 98, 1024]> linear_4_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_4_cast_fp16")];
+            tensor<int32, [4]> var_493 = const()[name = string("op_493"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_1_cast_fp16 = reshape(shape = var_493, x = linear_4_cast_fp16)[name = string("k_1_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15133184))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16181824))))[name = string("encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16183936)))];
+            tensor<fp16, [1, 98, 1024]> linear_5_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_5_cast_fp16")];
+            tensor<int32, [4]> var_498 = const()[name = string("op_498"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_1_cast_fp16 = reshape(shape = var_498, x = linear_5_cast_fp16)[name = string("v_1_cast_fp16")];
+            tensor<int32, [4]> value_9_perm_0 = const()[name = string("value_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16186048)))];
+            tensor<fp16, [1, 56, 8, 128]> var_511_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = string("op_511_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16188160)))];
+            tensor<fp16, [1, 56, 8, 128]> var_513_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = string("op_513_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_1_perm_0 = const()[name = string("q_with_bias_v_1_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_7_transpose_x_0 = const()[name = string("x_7_transpose_x_0"), val = bool(false)];
+            bool x_7_transpose_y_0 = const()[name = string("x_7_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_515_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16190272))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16390016))))[name = string("op_515_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_513_cast_fp16)[name = string("transpose_362")];
+            tensor<fp16, [1, 8, 56, 195]> x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = op_515_to_fp16_quantized)[name = string("x_7_cast_fp16")];
+            tensor<int32, [8]> x_9_pad_0 = const()[name = string("x_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_9_mode_0 = const()[name = string("x_9_mode_0"), val = string("constant")];
+            fp16 const_79_to_fp16 = const()[name = string("const_79_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_9_cast_fp16 = pad(constant_val = const_79_to_fp16, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7_cast_fp16)[name = string("x_9_cast_fp16")];
+            tensor<int32, [4]> var_523 = const()[name = string("op_523"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_11_cast_fp16 = reshape(shape = var_523, x = x_9_cast_fp16)[name = string("x_11_cast_fp16")];
+            tensor<int32, [4]> var_527_begin_0 = const()[name = string("op_527_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_527_end_0 = const()[name = string("op_527_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_527_end_mask_0 = const()[name = string("op_527_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_527_cast_fp16 = slice_by_index(begin = var_527_begin_0, end = var_527_end_0, end_mask = var_527_end_mask_0, x = x_11_cast_fp16)[name = string("op_527_cast_fp16")];
+            tensor<int32, [4]> var_528 = const()[name = string("op_528"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_1_cast_fp16 = reshape(shape = var_528, x = var_527_cast_fp16)[name = string("matrix_bd_1_cast_fp16")];
+            bool matrix_ac_1_transpose_x_0 = const()[name = string("matrix_ac_1_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_1_transpose_y_0 = const()[name = string("matrix_ac_1_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_97 = transpose(perm = transpose_97_perm_0, x = k_1_cast_fp16)[name = string("transpose_360")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_96 = transpose(perm = transpose_96_perm_0, x = var_511_cast_fp16)[name = string("transpose_361")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_96, y = transpose_97)[name = string("matrix_ac_1_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_3_begin_0 = const()[name = string("matrix_bd_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_3_end_0 = const()[name = string("matrix_bd_3_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_3_end_mask_0 = const()[name = string("matrix_bd_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1_cast_fp16)[name = string("matrix_bd_3_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_537_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = string("op_537_cast_fp16")];
+            fp16 _inversed_scores_1_y_0_to_fp16 = const()[name = string("_inversed_scores_1_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_1_cast_fp16 = mul(x = var_537_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = string("_inversed_scores_1_cast_fp16")];
+            tensor<int32, [1]> mask_11_axes_0 = const()[name = string("mask_11_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 56, 98]> mask_11 = expand_dims(axes = mask_11_axes_0, x = mask_9)[name = string("mask_11")];
+            fp16 var_44_to_fp16 = const()[name = string("op_44_to_fp16"), val = fp16(-0x1.388p+13)];
+            tensor<fp16, [1, 8, 56, 98]> scores_3_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_11)[name = string("scores_3_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_543_cast_fp16 = softmax(axis = var_58, x = scores_3_cast_fp16)[name = string("op_543_cast_fp16")];
+            fp16 var_43_to_fp16 = const()[name = string("op_43_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 98]> input_41_cast_fp16 = select(a = var_43_to_fp16, b = var_543_cast_fp16, cond = mask_11)[name = string("input_41_cast_fp16")];
+            bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)];
+            bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_1_cast_fp16)[name = string("transpose_359")];
+            tensor<fp16, [1, 8, 56, 128]> x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41_cast_fp16, y = value_9_cast_fp16)[name = string("x_13_cast_fp16")];
+            tensor<int32, [4]> var_547_perm_0 = const()[name = string("op_547_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_548 = const()[name = string("op_548"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_547_cast_fp16 = transpose(perm = var_547_perm_0, x = x_13_cast_fp16)[name = string("transpose_358")];
+            tensor<fp16, [1, 56, 1024]> input_43_cast_fp16 = reshape(shape = var_548, x = var_547_cast_fp16)[name = string("input_43_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16390528))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17439168))))[name = string("encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17441280)))];
+            tensor<fp16, [1, 56, 1024]> linear_7_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_7_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = string("input_47_cast_fp16")];
+            tensor<int32, [1]> x_17_axes_0 = const()[name = string("x_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17443392)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17445504)))];
+            tensor<fp16, [1, 56, 1024]> x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = string("x_17_cast_fp16")];
+            tensor<int32, [3]> input_49_perm_0 = const()[name = string("input_49_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_51_pad_type_0 = const()[name = string("input_51_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_51_strides_0 = const()[name = string("input_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_51_pad_0 = const()[name = string("input_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_51_dilations_0 = const()[name = string("input_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_51_groups_0 = const()[name = string("input_51_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17447616))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19544832))))[name = string("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = string("transpose_357")];
+            tensor<fp16, [1, 2048, 56]> input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("input_51_cast_fp16")];
+            int32 x_19_split_num_splits_0 = const()[name = string("x_19_split_num_splits_0"), val = int32(2)];
+            int32 x_19_split_axis_0 = const()[name = string("x_19_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_19_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51_cast_fp16)[name = string("x_19_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = string("x_19_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = string("x_19_cast_fp16")];
+            tensor<int32, [1]> var_574_axes_0 = const()[name = string("op_574_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 56]> var_574 = expand_dims(axes = var_574_axes_0, x = pad_mask)[name = string("op_574")];
+            tensor<fp16, [1, 1024, 56]> input_53_cast_fp16 = select(a = var_43_to_fp16, b = x_19_cast_fp16, cond = var_574)[name = string("input_53_cast_fp16")];
+            bool new_x_3_interleave_0 = const()[name = string("new_x_3_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_3_cast_fp16 = concat(axis = var_58, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = string("new_x_3_cast_fp16")];
+            tensor<int32, [3]> var_587_begin_0 = const()[name = string("op_587_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_587_end_0 = const()[name = string("op_587_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_587_end_mask_0 = const()[name = string("op_587_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_587_cast_fp16 = slice_by_index(begin = var_587_begin_0, end = var_587_end_0, end_mask = var_587_end_mask_0, x = new_x_3_cast_fp16)[name = string("op_587_cast_fp16")];
+            string x_21_pad_type_0 = const()[name = string("x_21_pad_type_0"), val = string("valid")];
+            int32 x_21_groups_0 = const()[name = string("x_21_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_21_strides_0 = const()[name = string("x_21_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_21_pad_0 = const()[name = string("x_21_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_21_dilations_0 = const()[name = string("x_21_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19548992))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19558272))))[name = string("encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_3_cast_fp16)[name = string("x_21_cast_fp16")];
+            tensor<int32, [3]> input_55_perm_0 = const()[name = string("input_55_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_23_axes_0 = const()[name = string("x_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19560384)))];
+            tensor<fp16, [1024]> encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19562496)))];
+            tensor<fp16, [1, 56, 1024]> input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = string("transpose_356")];
+            tensor<fp16, [1, 56, 1024]> x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = string("x_23_cast_fp16")];
+            tensor<int32, [3]> input_57_perm_0 = const()[name = string("input_57_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = string("transpose_355")];
+            tensor<fp16, [1, 1024, 56]> input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = string("input_59_cast_fp16")];
+            string x_25_pad_type_0 = const()[name = string("x_25_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_25_strides_0 = const()[name = string("x_25_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_25_pad_0 = const()[name = string("x_25_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_25_dilations_0 = const()[name = string("x_25_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_25_groups_0 = const()[name = string("x_25_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19564608))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20613248))))[name = string("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("x_25_cast_fp16")];
+            tensor<int32, [3]> input_61_perm_0 = const()[name = string("input_61_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = string("transpose_354")];
+            tensor<fp16, [1, 56, 1024]> input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = string("input_63_cast_fp16")];
+            tensor<int32, [1]> input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20615360)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20617472)))];
+            tensor<fp16, [1, 56, 1024]> input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20619584))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24813952))))[name = string("encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_0_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24822208)))];
+            tensor<fp16, [1, 56, 4096]> linear_8_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_8_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = string("input_69_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24830464))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29024832))))[name = string("encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29026944)))];
+            tensor<fp16, [1, 56, 1024]> linear_9_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_9_cast_fp16")];
+            fp16 var_630_to_fp16 = const()[name = string("op_630_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_631_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_630_to_fp16)[name = string("op_631_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_631_cast_fp16)[name = string("input_75_cast_fp16")];
+            tensor<int32, [1]> input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29029056)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29031168)))];
+            tensor<fp16, [1, 56, 1024]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")];
+            tensor<int32, [4]> cache_5_begin_0 = const()[name = string("cache_5_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_5_end_0 = const()[name = string("cache_5_end_0"), val = tensor<int32, [4]>([2, 1, 42, 1024])];
+            tensor<bool, [4]> cache_5_end_mask_0 = const()[name = string("cache_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_5_squeeze_mask_0 = const()[name = string("cache_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_5_cast_fp16 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_5_cast_fp16")];
+            tensor<int32, [4]> cache_7_begin_0 = const()[name = string("cache_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_7_end_0 = const()[name = string("cache_7_end_0"), val = tensor<int32, [4]>([2, 1, 1024, 8])];
+            tensor<bool, [4]> cache_7_end_mask_0 = const()[name = string("cache_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_7_squeeze_mask_0 = const()[name = string("cache_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_7_cast_fp16 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_7_cast_fp16")];
+            tensor<int32, [1]> input_79_axes_0 = const()[name = string("input_79_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29033280)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29035392)))];
+            tensor<fp16, [1, 56, 1024]> input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = string("input_79_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29037504))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33231872))))[name = string("encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_1_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33240128)))];
+            tensor<fp16, [1, 56, 4096]> linear_10_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_10_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = string("input_83_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33248384))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37442752))))[name = string("encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37444864)))];
+            tensor<fp16, [1, 56, 1024]> linear_11_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_11_cast_fp16")];
+            fp16 var_667_to_fp16 = const()[name = string("op_667_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_668_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_667_to_fp16)[name = string("op_668_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_668_cast_fp16)[name = string("input_89_cast_fp16")];
+            tensor<int32, [1]> key_3_axes_0 = const()[name = string("key_3_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37446976)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37449088)))];
+            tensor<fp16, [1, 56, 1024]> key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = string("key_3_cast_fp16")];
+            bool input_91_interleave_0 = const()[name = string("input_91_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_91_cast_fp16 = concat(axis = var_67, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = string("input_91_cast_fp16")];
+            bool var_696_interleave_0 = const()[name = string("op_696_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_696_cast_fp16 = concat(axis = var_67, interleave = var_696_interleave_0, values = key_3_cast_fp16)[name = string("op_696_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37451200))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38499840))))[name = string("encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38501952)))];
+            tensor<fp16, [1, 56, 1024]> linear_12_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized, x = key_3_cast_fp16)[name = string("linear_12_cast_fp16")];
+            tensor<int32, [4]> var_701 = const()[name = string("op_701"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_7_cast_fp16 = reshape(shape = var_701, x = linear_12_cast_fp16)[name = string("q_7_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38504064))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39552704))))[name = string("encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39554816)))];
+            tensor<fp16, [1, 98, 1024]> linear_13_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_13_cast_fp16")];
+            tensor<int32, [4]> var_706 = const()[name = string("op_706"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_5_cast_fp16 = reshape(shape = var_706, x = linear_13_cast_fp16)[name = string("k_5_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39556928))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40605568))))[name = string("encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40607680)))];
+            tensor<fp16, [1, 98, 1024]> linear_14_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_14_cast_fp16")];
+            tensor<int32, [4]> var_711 = const()[name = string("op_711"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_3_cast_fp16 = reshape(shape = var_711, x = linear_14_cast_fp16)[name = string("v_3_cast_fp16")];
+            tensor<int32, [4]> value_11_perm_0 = const()[name = string("value_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40609792)))];
+            tensor<fp16, [1, 56, 8, 128]> var_724_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = string("op_724_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40611904)))];
+            tensor<fp16, [1, 56, 8, 128]> var_726_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = string("op_726_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_3_perm_0 = const()[name = string("q_with_bias_v_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_33_transpose_x_0 = const()[name = string("x_33_transpose_x_0"), val = bool(false)];
+            bool x_33_transpose_y_0 = const()[name = string("x_33_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_728_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40614016))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40813760))))[name = string("op_728_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_726_cast_fp16)[name = string("transpose_353")];
+            tensor<fp16, [1, 8, 56, 195]> x_33_cast_fp16 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = op_728_to_fp16_quantized)[name = string("x_33_cast_fp16")];
+            tensor<int32, [8]> x_35_pad_0 = const()[name = string("x_35_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_35_mode_0 = const()[name = string("x_35_mode_0"), val = string("constant")];
+            fp16 const_92_to_fp16 = const()[name = string("const_92_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_35_cast_fp16 = pad(constant_val = const_92_to_fp16, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33_cast_fp16)[name = string("x_35_cast_fp16")];
+            tensor<int32, [4]> var_736 = const()[name = string("op_736"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_37_cast_fp16 = reshape(shape = var_736, x = x_35_cast_fp16)[name = string("x_37_cast_fp16")];
+            tensor<int32, [4]> var_740_begin_0 = const()[name = string("op_740_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_740_end_0 = const()[name = string("op_740_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_740_end_mask_0 = const()[name = string("op_740_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_740_cast_fp16 = slice_by_index(begin = var_740_begin_0, end = var_740_end_0, end_mask = var_740_end_mask_0, x = x_37_cast_fp16)[name = string("op_740_cast_fp16")];
+            tensor<int32, [4]> var_741 = const()[name = string("op_741"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_5_cast_fp16 = reshape(shape = var_741, x = var_740_cast_fp16)[name = string("matrix_bd_5_cast_fp16")];
+            bool matrix_ac_3_transpose_x_0 = const()[name = string("matrix_ac_3_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_3_transpose_y_0 = const()[name = string("matrix_ac_3_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_99 = transpose(perm = transpose_99_perm_0, x = k_5_cast_fp16)[name = string("transpose_351")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_98 = transpose(perm = transpose_98_perm_0, x = var_724_cast_fp16)[name = string("transpose_352")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_98, y = transpose_99)[name = string("matrix_ac_3_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_7_begin_0 = const()[name = string("matrix_bd_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_7_end_0 = const()[name = string("matrix_bd_7_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_7_end_mask_0 = const()[name = string("matrix_bd_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5_cast_fp16)[name = string("matrix_bd_7_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_750_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = string("op_750_cast_fp16")];
+            fp16 _inversed_scores_5_y_0_to_fp16 = const()[name = string("_inversed_scores_5_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_5_cast_fp16 = mul(x = var_750_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = string("_inversed_scores_5_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_7_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_11)[name = string("scores_7_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_756_cast_fp16 = softmax(axis = var_58, x = scores_7_cast_fp16)[name = string("op_756_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_93_cast_fp16 = select(a = var_43_to_fp16, b = var_756_cast_fp16, cond = mask_11)[name = string("input_93_cast_fp16")];
+            bool x_39_transpose_x_0 = const()[name = string("x_39_transpose_x_0"), val = bool(false)];
+            bool x_39_transpose_y_0 = const()[name = string("x_39_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_3_cast_fp16)[name = string("transpose_350")];
+            tensor<fp16, [1, 8, 56, 128]> x_39_cast_fp16 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93_cast_fp16, y = value_11_cast_fp16)[name = string("x_39_cast_fp16")];
+            tensor<int32, [4]> var_760_perm_0 = const()[name = string("op_760_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_761 = const()[name = string("op_761"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_760_cast_fp16 = transpose(perm = var_760_perm_0, x = x_39_cast_fp16)[name = string("transpose_349")];
+            tensor<fp16, [1, 56, 1024]> input_95_cast_fp16 = reshape(shape = var_761, x = var_760_cast_fp16)[name = string("input_95_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40814272))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41862912))))[name = string("encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41865024)))];
+            tensor<fp16, [1, 56, 1024]> linear_16_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_16_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = string("input_99_cast_fp16")];
+            tensor<int32, [1]> x_43_axes_0 = const()[name = string("x_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41867136)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41869248)))];
+            tensor<fp16, [1, 56, 1024]> x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = string("x_43_cast_fp16")];
+            tensor<int32, [3]> input_101_perm_0 = const()[name = string("input_101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_103_pad_type_0 = const()[name = string("input_103_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_103_strides_0 = const()[name = string("input_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_103_pad_0 = const()[name = string("input_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_103_dilations_0 = const()[name = string("input_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_103_groups_0 = const()[name = string("input_103_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41871360))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43968576))))[name = string("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = string("transpose_348")];
+            tensor<fp16, [1, 2048, 56]> input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")];
+            int32 x_45_split_num_splits_0 = const()[name = string("x_45_split_num_splits_0"), val = int32(2)];
+            int32 x_45_split_axis_0 = const()[name = string("x_45_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_45_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_45_split_cast_fp16_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103_cast_fp16)[name = string("x_45_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = string("x_45_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = string("x_45_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_105_cast_fp16 = select(a = var_43_to_fp16, b = x_45_cast_fp16, cond = var_574)[name = string("input_105_cast_fp16")];
+            bool new_x_7_interleave_0 = const()[name = string("new_x_7_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_7_cast_fp16 = concat(axis = var_58, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = string("new_x_7_cast_fp16")];
+            tensor<int32, [3]> var_800_begin_0 = const()[name = string("op_800_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_800_end_0 = const()[name = string("op_800_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_800_end_mask_0 = const()[name = string("op_800_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_800_cast_fp16 = slice_by_index(begin = var_800_begin_0, end = var_800_end_0, end_mask = var_800_end_mask_0, x = new_x_7_cast_fp16)[name = string("op_800_cast_fp16")];
+            string x_47_pad_type_0 = const()[name = string("x_47_pad_type_0"), val = string("valid")];
+            int32 x_47_groups_0 = const()[name = string("x_47_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_47_strides_0 = const()[name = string("x_47_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_47_pad_0 = const()[name = string("x_47_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_47_dilations_0 = const()[name = string("x_47_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43972736))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43982016))))[name = string("encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_7_cast_fp16)[name = string("x_47_cast_fp16")];
+            tensor<int32, [3]> input_107_perm_0 = const()[name = string("input_107_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_49_axes_0 = const()[name = string("x_49_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43984128)))];
+            tensor<fp16, [1024]> encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43986240)))];
+            tensor<fp16, [1, 56, 1024]> input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = string("transpose_347")];
+            tensor<fp16, [1, 56, 1024]> x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = string("x_49_cast_fp16")];
+            tensor<int32, [3]> input_109_perm_0 = const()[name = string("input_109_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = string("transpose_346")];
+            tensor<fp16, [1, 1024, 56]> input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = string("input_111_cast_fp16")];
+            string x_51_pad_type_0 = const()[name = string("x_51_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_51_strides_0 = const()[name = string("x_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_51_pad_0 = const()[name = string("x_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_51_dilations_0 = const()[name = string("x_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_51_groups_0 = const()[name = string("x_51_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43988352))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45036992))))[name = string("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("x_51_cast_fp16")];
+            tensor<int32, [3]> input_113_perm_0 = const()[name = string("input_113_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = string("transpose_345")];
+            tensor<fp16, [1, 56, 1024]> input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = string("input_115_cast_fp16")];
+            tensor<int32, [1]> input_117_axes_0 = const()[name = string("input_117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45039104)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45041216)))];
+            tensor<fp16, [1, 56, 1024]> input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = string("input_117_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45043328))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49237696))))[name = string("encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_1_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49245952)))];
+            tensor<fp16, [1, 56, 4096]> linear_17_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_17_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = string("input_121_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49254208))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53448576))))[name = string("encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53450688)))];
+            tensor<fp16, [1, 56, 1024]> linear_18_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_18_cast_fp16")];
+            fp16 var_843_to_fp16 = const()[name = string("op_843_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_844_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_843_to_fp16)[name = string("op_844_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_844_cast_fp16)[name = string("input_127_cast_fp16")];
+            tensor<int32, [1]> input_129_axes_0 = const()[name = string("input_129_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53452800)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53454912)))];
+            tensor<fp16, [1, 56, 1024]> input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = string("input_129_cast_fp16")];
+            tensor<int32, [4]> cache_9_begin_0 = const()[name = string("cache_9_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_9_end_0 = const()[name = string("cache_9_end_0"), val = tensor<int32, [4]>([3, 1, 42, 1024])];
+            tensor<bool, [4]> cache_9_end_mask_0 = const()[name = string("cache_9_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_9_squeeze_mask_0 = const()[name = string("cache_9_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_9_cast_fp16 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_9_cast_fp16")];
+            tensor<int32, [4]> cache_11_begin_0 = const()[name = string("cache_11_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_11_end_0 = const()[name = string("cache_11_end_0"), val = tensor<int32, [4]>([3, 1, 1024, 8])];
+            tensor<bool, [4]> cache_11_end_mask_0 = const()[name = string("cache_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_11_squeeze_mask_0 = const()[name = string("cache_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_11_cast_fp16 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_11_cast_fp16")];
+            tensor<int32, [1]> input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53457024)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53459136)))];
+            tensor<fp16, [1, 56, 1024]> input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53461248))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57655616))))[name = string("encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_2_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57663872)))];
+            tensor<fp16, [1, 56, 4096]> linear_19_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_19_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = string("input_135_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57672128))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61866496))))[name = string("encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61868608)))];
+            tensor<fp16, [1, 56, 1024]> linear_20_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_20_cast_fp16")];
+            fp16 var_880_to_fp16 = const()[name = string("op_880_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_881_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_880_to_fp16)[name = string("op_881_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_881_cast_fp16)[name = string("input_141_cast_fp16")];
+            tensor<int32, [1]> key_5_axes_0 = const()[name = string("key_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61870720)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61872832)))];
+            tensor<fp16, [1, 56, 1024]> key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = string("key_5_cast_fp16")];
+            bool input_143_interleave_0 = const()[name = string("input_143_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_143_cast_fp16 = concat(axis = var_67, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = string("input_143_cast_fp16")];
+            bool var_909_interleave_0 = const()[name = string("op_909_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_909_cast_fp16 = concat(axis = var_67, interleave = var_909_interleave_0, values = key_5_cast_fp16)[name = string("op_909_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61874944))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62923584))))[name = string("encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62925696)))];
+            tensor<fp16, [1, 56, 1024]> linear_21_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized, x = key_5_cast_fp16)[name = string("linear_21_cast_fp16")];
+            tensor<int32, [4]> var_914 = const()[name = string("op_914"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_13_cast_fp16 = reshape(shape = var_914, x = linear_21_cast_fp16)[name = string("q_13_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62927808))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63976448))))[name = string("encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63978560)))];
+            tensor<fp16, [1, 98, 1024]> linear_22_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_22_cast_fp16")];
+            tensor<int32, [4]> var_919 = const()[name = string("op_919"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_9_cast_fp16 = reshape(shape = var_919, x = linear_22_cast_fp16)[name = string("k_9_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63980672))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65029312))))[name = string("encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65031424)))];
+            tensor<fp16, [1, 98, 1024]> linear_23_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_23_cast_fp16")];
+            tensor<int32, [4]> var_924 = const()[name = string("op_924"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_5_cast_fp16 = reshape(shape = var_924, x = linear_23_cast_fp16)[name = string("v_5_cast_fp16")];
+            tensor<int32, [4]> value_13_perm_0 = const()[name = string("value_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65033536)))];
+            tensor<fp16, [1, 56, 8, 128]> var_937_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = string("op_937_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65035648)))];
+            tensor<fp16, [1, 56, 8, 128]> var_939_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = string("op_939_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_5_perm_0 = const()[name = string("q_with_bias_v_5_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_59_transpose_x_0 = const()[name = string("x_59_transpose_x_0"), val = bool(false)];
+            bool x_59_transpose_y_0 = const()[name = string("x_59_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_941_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65037760))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65237504))))[name = string("op_941_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_939_cast_fp16)[name = string("transpose_344")];
+            tensor<fp16, [1, 8, 56, 195]> x_59_cast_fp16 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = op_941_to_fp16_quantized)[name = string("x_59_cast_fp16")];
+            tensor<int32, [8]> x_61_pad_0 = const()[name = string("x_61_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_61_mode_0 = const()[name = string("x_61_mode_0"), val = string("constant")];
+            fp16 const_105_to_fp16 = const()[name = string("const_105_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_61_cast_fp16 = pad(constant_val = const_105_to_fp16, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59_cast_fp16)[name = string("x_61_cast_fp16")];
+            tensor<int32, [4]> var_949 = const()[name = string("op_949"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_63_cast_fp16 = reshape(shape = var_949, x = x_61_cast_fp16)[name = string("x_63_cast_fp16")];
+            tensor<int32, [4]> var_953_begin_0 = const()[name = string("op_953_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_953_end_0 = const()[name = string("op_953_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_953_end_mask_0 = const()[name = string("op_953_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_953_cast_fp16 = slice_by_index(begin = var_953_begin_0, end = var_953_end_0, end_mask = var_953_end_mask_0, x = x_63_cast_fp16)[name = string("op_953_cast_fp16")];
+            tensor<int32, [4]> var_954 = const()[name = string("op_954"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_9_cast_fp16 = reshape(shape = var_954, x = var_953_cast_fp16)[name = string("matrix_bd_9_cast_fp16")];
+            bool matrix_ac_5_transpose_x_0 = const()[name = string("matrix_ac_5_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_5_transpose_y_0 = const()[name = string("matrix_ac_5_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_101 = transpose(perm = transpose_101_perm_0, x = k_9_cast_fp16)[name = string("transpose_342")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_100 = transpose(perm = transpose_100_perm_0, x = var_937_cast_fp16)[name = string("transpose_343")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_100, y = transpose_101)[name = string("matrix_ac_5_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_11_begin_0 = const()[name = string("matrix_bd_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_11_end_0 = const()[name = string("matrix_bd_11_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_11_end_mask_0 = const()[name = string("matrix_bd_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9_cast_fp16)[name = string("matrix_bd_11_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_963_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = string("op_963_cast_fp16")];
+            fp16 _inversed_scores_9_y_0_to_fp16 = const()[name = string("_inversed_scores_9_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_9_cast_fp16 = mul(x = var_963_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = string("_inversed_scores_9_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_11_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_11)[name = string("scores_11_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_969_cast_fp16 = softmax(axis = var_58, x = scores_11_cast_fp16)[name = string("op_969_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_145_cast_fp16 = select(a = var_43_to_fp16, b = var_969_cast_fp16, cond = mask_11)[name = string("input_145_cast_fp16")];
+            bool x_65_transpose_x_0 = const()[name = string("x_65_transpose_x_0"), val = bool(false)];
+            bool x_65_transpose_y_0 = const()[name = string("x_65_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_5_cast_fp16)[name = string("transpose_341")];
+            tensor<fp16, [1, 8, 56, 128]> x_65_cast_fp16 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145_cast_fp16, y = value_13_cast_fp16)[name = string("x_65_cast_fp16")];
+            tensor<int32, [4]> var_973_perm_0 = const()[name = string("op_973_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_974 = const()[name = string("op_974"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_973_cast_fp16 = transpose(perm = var_973_perm_0, x = x_65_cast_fp16)[name = string("transpose_340")];
+            tensor<fp16, [1, 56, 1024]> input_147_cast_fp16 = reshape(shape = var_974, x = var_973_cast_fp16)[name = string("input_147_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65238016))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66024512))))[name = string("encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66024704)))];
+            tensor<fp16, [1, 56, 1024]> linear_25_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = string("linear_25_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = string("input_151_cast_fp16")];
+            tensor<int32, [1]> x_69_axes_0 = const()[name = string("x_69_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66026816)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66028928)))];
+            tensor<fp16, [1, 56, 1024]> x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = string("x_69_cast_fp16")];
+            tensor<int32, [3]> input_153_perm_0 = const()[name = string("input_153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_155_pad_type_0 = const()[name = string("input_155_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_155_strides_0 = const()[name = string("input_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_155_pad_0 = const()[name = string("input_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_155_dilations_0 = const()[name = string("input_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_155_groups_0 = const()[name = string("input_155_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66031040))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68128256))))[name = string("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = string("transpose_339")];
+            tensor<fp16, [1, 2048, 56]> input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")];
+            int32 x_71_split_num_splits_0 = const()[name = string("x_71_split_num_splits_0"), val = int32(2)];
+            int32 x_71_split_axis_0 = const()[name = string("x_71_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_71_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_71_split_cast_fp16_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155_cast_fp16)[name = string("x_71_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = string("x_71_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = string("x_71_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_157_cast_fp16 = select(a = var_43_to_fp16, b = x_71_cast_fp16, cond = var_574)[name = string("input_157_cast_fp16")];
+            bool new_x_11_interleave_0 = const()[name = string("new_x_11_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_11_cast_fp16 = concat(axis = var_58, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = string("new_x_11_cast_fp16")];
+            tensor<int32, [3]> var_1013_begin_0 = const()[name = string("op_1013_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_1013_end_0 = const()[name = string("op_1013_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_1013_end_mask_0 = const()[name = string("op_1013_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1013_cast_fp16 = slice_by_index(begin = var_1013_begin_0, end = var_1013_end_0, end_mask = var_1013_end_mask_0, x = new_x_11_cast_fp16)[name = string("op_1013_cast_fp16")];
+            string x_73_pad_type_0 = const()[name = string("x_73_pad_type_0"), val = string("valid")];
+            int32 x_73_groups_0 = const()[name = string("x_73_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_73_strides_0 = const()[name = string("x_73_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_73_pad_0 = const()[name = string("x_73_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_73_dilations_0 = const()[name = string("x_73_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68132416))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68141696))))[name = string("encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_73_cast_fp16 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_11_cast_fp16)[name = string("x_73_cast_fp16")];
+            tensor<int32, [3]> input_159_perm_0 = const()[name = string("input_159_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_75_axes_0 = const()[name = string("x_75_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68143808)))];
+            tensor<fp16, [1024]> encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68145920)))];
+            tensor<fp16, [1, 56, 1024]> input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = string("transpose_338")];
+            tensor<fp16, [1, 56, 1024]> x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = string("x_75_cast_fp16")];
+            tensor<int32, [3]> input_161_perm_0 = const()[name = string("input_161_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = string("transpose_337")];
+            tensor<fp16, [1, 1024, 56]> input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = string("input_163_cast_fp16")];
+            string x_77_pad_type_0 = const()[name = string("x_77_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_77_strides_0 = const()[name = string("x_77_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_77_pad_0 = const()[name = string("x_77_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_77_dilations_0 = const()[name = string("x_77_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_77_groups_0 = const()[name = string("x_77_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68148032))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69196672))))[name = string("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("x_77_cast_fp16")];
+            tensor<int32, [3]> input_165_perm_0 = const()[name = string("input_165_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = string("transpose_336")];
+            tensor<fp16, [1, 56, 1024]> input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = string("input_167_cast_fp16")];
+            tensor<int32, [1]> input_169_axes_0 = const()[name = string("input_169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69198784)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69200896)))];
+            tensor<fp16, [1, 56, 1024]> input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = string("input_169_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69203008))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72348800))))[name = string("encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_2_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72348992)))];
+            tensor<fp16, [1, 56, 4096]> linear_26_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = string("linear_26_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = string("input_173_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72357248))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75503040))))[name = string("encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_2_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75503232)))];
+            tensor<fp16, [1, 56, 1024]> linear_27_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized, x = input_173_cast_fp16)[name = string("linear_27_cast_fp16")];
+            fp16 var_1056_to_fp16 = const()[name = string("op_1056_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1057_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1056_to_fp16)[name = string("op_1057_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_1057_cast_fp16)[name = string("input_179_cast_fp16")];
+            tensor<int32, [1]> input_181_axes_0 = const()[name = string("input_181_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75505344)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75507456)))];
+            tensor<fp16, [1, 56, 1024]> input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = string("input_181_cast_fp16")];
+            tensor<int32, [4]> cache_13_begin_0 = const()[name = string("cache_13_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_13_end_0 = const()[name = string("cache_13_end_0"), val = tensor<int32, [4]>([4, 1, 42, 1024])];
+            tensor<bool, [4]> cache_13_end_mask_0 = const()[name = string("cache_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_13_squeeze_mask_0 = const()[name = string("cache_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_13_cast_fp16 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_13_cast_fp16")];
+            tensor<int32, [4]> cache_15_begin_0 = const()[name = string("cache_15_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_15_end_0 = const()[name = string("cache_15_end_0"), val = tensor<int32, [4]>([4, 1, 1024, 8])];
+            tensor<bool, [4]> cache_15_end_mask_0 = const()[name = string("cache_15_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_15_squeeze_mask_0 = const()[name = string("cache_15_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_15_cast_fp16 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_15_cast_fp16")];
+            tensor<int32, [1]> input_183_axes_0 = const()[name = string("input_183_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75509568)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75511680)))];
+            tensor<fp16, [1, 56, 1024]> input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = string("input_183_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75513792))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78659584))))[name = string("encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_3_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78659776)))];
+            tensor<fp16, [1, 56, 4096]> linear_28_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = string("linear_28_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = string("input_187_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78668032))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81813824))))[name = string("encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81814016)))];
+            tensor<fp16, [1, 56, 1024]> linear_29_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = string("linear_29_cast_fp16")];
+            fp16 var_1093_to_fp16 = const()[name = string("op_1093_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1094_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_1093_to_fp16)[name = string("op_1094_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_1094_cast_fp16)[name = string("input_193_cast_fp16")];
+            tensor<int32, [1]> key_7_axes_0 = const()[name = string("key_7_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81816128)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81818240)))];
+            tensor<fp16, [1, 56, 1024]> key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = string("key_7_cast_fp16")];
+            bool input_195_interleave_0 = const()[name = string("input_195_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_195_cast_fp16 = concat(axis = var_67, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = string("input_195_cast_fp16")];
+            bool var_1122_interleave_0 = const()[name = string("op_1122_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_1122_cast_fp16 = concat(axis = var_67, interleave = var_1122_interleave_0, values = key_7_cast_fp16)[name = string("op_1122_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81820352))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82606848))))[name = string("encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82607040)))];
+            tensor<fp16, [1, 56, 1024]> linear_30_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized, x = key_7_cast_fp16)[name = string("linear_30_cast_fp16")];
+            tensor<int32, [4]> var_1127 = const()[name = string("op_1127"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_19_cast_fp16 = reshape(shape = var_1127, x = linear_30_cast_fp16)[name = string("q_19_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82609152))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83395648))))[name = string("encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83395840)))];
+            tensor<fp16, [1, 98, 1024]> linear_31_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_31_cast_fp16")];
+            tensor<int32, [4]> var_1132 = const()[name = string("op_1132"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_13_cast_fp16 = reshape(shape = var_1132, x = linear_31_cast_fp16)[name = string("k_13_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83397952))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84184448))))[name = string("encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84184640)))];
+            tensor<fp16, [1, 98, 1024]> linear_32_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_32_cast_fp16")];
+            tensor<int32, [4]> var_1137 = const()[name = string("op_1137"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_7_cast_fp16 = reshape(shape = var_1137, x = linear_32_cast_fp16)[name = string("v_7_cast_fp16")];
+            tensor<int32, [4]> value_15_perm_0 = const()[name = string("value_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84186752)))];
+            tensor<fp16, [1, 56, 8, 128]> var_1150_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = string("op_1150_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84188864)))];
+            tensor<fp16, [1, 56, 8, 128]> var_1152_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = string("op_1152_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_7_perm_0 = const()[name = string("q_with_bias_v_7_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_85_transpose_x_0 = const()[name = string("x_85_transpose_x_0"), val = bool(false)];
+            bool x_85_transpose_y_0 = const()[name = string("x_85_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_1154_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84190976))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84390720))))[name = string("op_1154_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_1152_cast_fp16)[name = string("transpose_335")];
+            tensor<fp16, [1, 8, 56, 195]> x_85_cast_fp16 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = op_1154_to_fp16_quantized)[name = string("x_85_cast_fp16")];
+            tensor<int32, [8]> x_87_pad_0 = const()[name = string("x_87_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("constant")];
+            fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_87_cast_fp16 = pad(constant_val = const_118_to_fp16, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85_cast_fp16)[name = string("x_87_cast_fp16")];
+            tensor<int32, [4]> var_1162 = const()[name = string("op_1162"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_89_cast_fp16 = reshape(shape = var_1162, x = x_87_cast_fp16)[name = string("x_89_cast_fp16")];
+            tensor<int32, [4]> var_1166_begin_0 = const()[name = string("op_1166_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1166_end_0 = const()[name = string("op_1166_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_1166_end_mask_0 = const()[name = string("op_1166_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_1166_cast_fp16 = slice_by_index(begin = var_1166_begin_0, end = var_1166_end_0, end_mask = var_1166_end_mask_0, x = x_89_cast_fp16)[name = string("op_1166_cast_fp16")];
+            tensor<int32, [4]> var_1167 = const()[name = string("op_1167"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_13_cast_fp16 = reshape(shape = var_1167, x = var_1166_cast_fp16)[name = string("matrix_bd_13_cast_fp16")];
+            bool matrix_ac_7_transpose_x_0 = const()[name = string("matrix_ac_7_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_7_transpose_y_0 = const()[name = string("matrix_ac_7_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_103 = transpose(perm = transpose_103_perm_0, x = k_13_cast_fp16)[name = string("transpose_333")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_102 = transpose(perm = transpose_102_perm_0, x = var_1150_cast_fp16)[name = string("transpose_334")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_102, y = transpose_103)[name = string("matrix_ac_7_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_15_begin_0 = const()[name = string("matrix_bd_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_15_end_0 = const()[name = string("matrix_bd_15_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_15_end_mask_0 = const()[name = string("matrix_bd_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13_cast_fp16)[name = string("matrix_bd_15_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_1176_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = string("op_1176_cast_fp16")];
+            fp16 _inversed_scores_13_y_0_to_fp16 = const()[name = string("_inversed_scores_13_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_13_cast_fp16 = mul(x = var_1176_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = string("_inversed_scores_13_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_15_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_11)[name = string("scores_15_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_1182_cast_fp16 = softmax(axis = var_58, x = scores_15_cast_fp16)[name = string("op_1182_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_197_cast_fp16 = select(a = var_43_to_fp16, b = var_1182_cast_fp16, cond = mask_11)[name = string("input_197_cast_fp16")];
+            bool x_91_transpose_x_0 = const()[name = string("x_91_transpose_x_0"), val = bool(false)];
+            bool x_91_transpose_y_0 = const()[name = string("x_91_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_7_cast_fp16)[name = string("transpose_332")];
+            tensor<fp16, [1, 8, 56, 128]> x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197_cast_fp16, y = value_15_cast_fp16)[name = string("x_91_cast_fp16")];
+            tensor<int32, [4]> var_1186_perm_0 = const()[name = string("op_1186_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1187 = const()[name = string("op_1187"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_1186_cast_fp16 = transpose(perm = var_1186_perm_0, x = x_91_cast_fp16)[name = string("transpose_331")];
+            tensor<fp16, [1, 56, 1024]> input_199_cast_fp16 = reshape(shape = var_1187, x = var_1186_cast_fp16)[name = string("input_199_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84391232))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85177728))))[name = string("encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85177920)))];
+            tensor<fp16, [1, 56, 1024]> linear_34_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = string("linear_34_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = string("input_203_cast_fp16")];
+            tensor<int32, [1]> x_95_axes_0 = const()[name = string("x_95_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85180032)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85182144)))];
+            tensor<fp16, [1, 56, 1024]> x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = string("x_95_cast_fp16")];
+            tensor<int32, [3]> input_205_perm_0 = const()[name = string("input_205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_207_pad_type_0 = const()[name = string("input_207_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_207_strides_0 = const()[name = string("input_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_207_pad_0 = const()[name = string("input_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_207_dilations_0 = const()[name = string("input_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_207_groups_0 = const()[name = string("input_207_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85184256))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87281472))))[name = string("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = string("transpose_330")];
+            tensor<fp16, [1, 2048, 56]> input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("input_207_cast_fp16")];
+            int32 x_97_split_num_splits_0 = const()[name = string("x_97_split_num_splits_0"), val = int32(2)];
+            int32 x_97_split_axis_0 = const()[name = string("x_97_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_97_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207_cast_fp16)[name = string("x_97_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = string("x_97_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = string("x_97_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_209_cast_fp16 = select(a = var_43_to_fp16, b = x_97_cast_fp16, cond = var_574)[name = string("input_209_cast_fp16")];
+            bool new_x_15_interleave_0 = const()[name = string("new_x_15_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_15_cast_fp16 = concat(axis = var_58, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = string("new_x_15_cast_fp16")];
+            tensor<int32, [3]> var_1226_begin_0 = const()[name = string("op_1226_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_1226_end_0 = const()[name = string("op_1226_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_1226_end_mask_0 = const()[name = string("op_1226_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1226_cast_fp16 = slice_by_index(begin = var_1226_begin_0, end = var_1226_end_0, end_mask = var_1226_end_mask_0, x = new_x_15_cast_fp16)[name = string("op_1226_cast_fp16")];
+            string x_99_pad_type_0 = const()[name = string("x_99_pad_type_0"), val = string("valid")];
+            int32 x_99_groups_0 = const()[name = string("x_99_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_99_strides_0 = const()[name = string("x_99_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_99_pad_0 = const()[name = string("x_99_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_99_dilations_0 = const()[name = string("x_99_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87285632))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87294912))))[name = string("encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_15_cast_fp16)[name = string("x_99_cast_fp16")];
+            tensor<int32, [3]> input_211_perm_0 = const()[name = string("input_211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_101_axes_0 = const()[name = string("x_101_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87297024)))];
+            tensor<fp16, [1024]> encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87299136)))];
+            tensor<fp16, [1, 56, 1024]> input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = string("transpose_329")];
+            tensor<fp16, [1, 56, 1024]> x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = string("x_101_cast_fp16")];
+            tensor<int32, [3]> input_213_perm_0 = const()[name = string("input_213_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = string("transpose_328")];
+            tensor<fp16, [1, 1024, 56]> input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = string("input_215_cast_fp16")];
+            string x_103_pad_type_0 = const()[name = string("x_103_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_103_strides_0 = const()[name = string("x_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_103_pad_0 = const()[name = string("x_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_103_dilations_0 = const()[name = string("x_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_103_groups_0 = const()[name = string("x_103_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87301248))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88349888))))[name = string("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("x_103_cast_fp16")];
+            tensor<int32, [3]> input_217_perm_0 = const()[name = string("input_217_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = string("transpose_327")];
+            tensor<fp16, [1, 56, 1024]> input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = string("input_219_cast_fp16")];
+            tensor<int32, [1]> input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88352000)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88354112)))];
+            tensor<fp16, [1, 56, 1024]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88356224))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91502016))))[name = string("encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_3_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91502208)))];
+            tensor<fp16, [1, 56, 4096]> linear_35_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = string("linear_35_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = string("input_225_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91510464))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94656256))))[name = string("encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94656448)))];
+            tensor<fp16, [1, 56, 1024]> linear_36_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = string("linear_36_cast_fp16")];
+            fp16 var_1269_to_fp16 = const()[name = string("op_1269_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1270_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1269_to_fp16)[name = string("op_1270_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1270_cast_fp16)[name = string("input_231_cast_fp16")];
+            tensor<int32, [1]> input_233_axes_0 = const()[name = string("input_233_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94658560)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94660672)))];
+            tensor<fp16, [1, 56, 1024]> input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = string("input_233_cast_fp16")];
+            tensor<int32, [4]> cache_17_begin_0 = const()[name = string("cache_17_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_17_end_0 = const()[name = string("cache_17_end_0"), val = tensor<int32, [4]>([5, 1, 42, 1024])];
+            tensor<bool, [4]> cache_17_end_mask_0 = const()[name = string("cache_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_17_squeeze_mask_0 = const()[name = string("cache_17_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_17_cast_fp16 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_17_cast_fp16")];
+            tensor<int32, [4]> cache_19_begin_0 = const()[name = string("cache_19_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_19_end_0 = const()[name = string("cache_19_end_0"), val = tensor<int32, [4]>([5, 1, 1024, 8])];
+            tensor<bool, [4]> cache_19_end_mask_0 = const()[name = string("cache_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_19_squeeze_mask_0 = const()[name = string("cache_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_19_cast_fp16 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_19_cast_fp16")];
+            tensor<int32, [1]> input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94662784)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94664896)))];
+            tensor<fp16, [1, 56, 1024]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94667008))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97812800))))[name = string("encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_4_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97812992)))];
+            tensor<fp16, [1, 56, 4096]> linear_37_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized, x = input_235_cast_fp16)[name = string("linear_37_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = string("input_239_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97821248))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100967040))))[name = string("encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100967232)))];
+            tensor<fp16, [1, 56, 1024]> linear_38_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = string("linear_38_cast_fp16")];
+            fp16 var_1306_to_fp16 = const()[name = string("op_1306_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1307_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1306_to_fp16)[name = string("op_1307_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1307_cast_fp16)[name = string("input_245_cast_fp16")];
+            tensor<int32, [1]> key_9_axes_0 = const()[name = string("key_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100969344)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100971456)))];
+            tensor<fp16, [1, 56, 1024]> key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = string("key_9_cast_fp16")];
+            bool input_247_interleave_0 = const()[name = string("input_247_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_247_cast_fp16 = concat(axis = var_67, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = string("input_247_cast_fp16")];
+            bool var_1335_interleave_0 = const()[name = string("op_1335_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_1335_cast_fp16 = concat(axis = var_67, interleave = var_1335_interleave_0, values = key_9_cast_fp16)[name = string("op_1335_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100973568))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101760064))))[name = string("encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101760256)))];
+            tensor<fp16, [1, 56, 1024]> linear_39_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized, x = key_9_cast_fp16)[name = string("linear_39_cast_fp16")];
+            tensor<int32, [4]> var_1340 = const()[name = string("op_1340"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_25_cast_fp16 = reshape(shape = var_1340, x = linear_39_cast_fp16)[name = string("q_25_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101762368))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102548864))))[name = string("encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102549056)))];
+            tensor<fp16, [1, 98, 1024]> linear_40_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_40_cast_fp16")];
+            tensor<int32, [4]> var_1345 = const()[name = string("op_1345"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_17_cast_fp16 = reshape(shape = var_1345, x = linear_40_cast_fp16)[name = string("k_17_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102551168))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103337664))))[name = string("encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103337856)))];
+            tensor<fp16, [1, 98, 1024]> linear_41_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_41_cast_fp16")];
+            tensor<int32, [4]> var_1350 = const()[name = string("op_1350"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_9_cast_fp16 = reshape(shape = var_1350, x = linear_41_cast_fp16)[name = string("v_9_cast_fp16")];
+            tensor<int32, [4]> value_17_perm_0 = const()[name = string("value_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103339968)))];
+            tensor<fp16, [1, 56, 8, 128]> var_1363_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = string("op_1363_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103342080)))];
+            tensor<fp16, [1, 56, 8, 128]> var_1365_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = string("op_1365_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_9_perm_0 = const()[name = string("q_with_bias_v_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_111_transpose_x_0 = const()[name = string("x_111_transpose_x_0"), val = bool(false)];
+            bool x_111_transpose_y_0 = const()[name = string("x_111_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_1367_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103344192))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103543936))))[name = string("op_1367_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1365_cast_fp16)[name = string("transpose_326")];
+            tensor<fp16, [1, 8, 56, 195]> x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = op_1367_to_fp16_quantized)[name = string("x_111_cast_fp16")];
+            tensor<int32, [8]> x_113_pad_0 = const()[name = string("x_113_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_113_mode_0 = const()[name = string("x_113_mode_0"), val = string("constant")];
+            fp16 const_131_to_fp16 = const()[name = string("const_131_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_113_cast_fp16 = pad(constant_val = const_131_to_fp16, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111_cast_fp16)[name = string("x_113_cast_fp16")];
+            tensor<int32, [4]> var_1375 = const()[name = string("op_1375"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_115_cast_fp16 = reshape(shape = var_1375, x = x_113_cast_fp16)[name = string("x_115_cast_fp16")];
+            tensor<int32, [4]> var_1379_begin_0 = const()[name = string("op_1379_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1379_end_0 = const()[name = string("op_1379_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_1379_end_mask_0 = const()[name = string("op_1379_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_1379_cast_fp16 = slice_by_index(begin = var_1379_begin_0, end = var_1379_end_0, end_mask = var_1379_end_mask_0, x = x_115_cast_fp16)[name = string("op_1379_cast_fp16")];
+            tensor<int32, [4]> var_1380 = const()[name = string("op_1380"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_17_cast_fp16 = reshape(shape = var_1380, x = var_1379_cast_fp16)[name = string("matrix_bd_17_cast_fp16")];
+            bool matrix_ac_9_transpose_x_0 = const()[name = string("matrix_ac_9_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_9_transpose_y_0 = const()[name = string("matrix_ac_9_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_105 = transpose(perm = transpose_105_perm_0, x = k_17_cast_fp16)[name = string("transpose_324")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1363_cast_fp16)[name = string("transpose_325")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_104, y = transpose_105)[name = string("matrix_ac_9_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_19_begin_0 = const()[name = string("matrix_bd_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_19_end_0 = const()[name = string("matrix_bd_19_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_19_end_mask_0 = const()[name = string("matrix_bd_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17_cast_fp16)[name = string("matrix_bd_19_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_1389_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = string("op_1389_cast_fp16")];
+            fp16 _inversed_scores_17_y_0_to_fp16 = const()[name = string("_inversed_scores_17_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_17_cast_fp16 = mul(x = var_1389_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = string("_inversed_scores_17_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_19_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_11)[name = string("scores_19_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_1395_cast_fp16 = softmax(axis = var_58, x = scores_19_cast_fp16)[name = string("op_1395_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_249_cast_fp16 = select(a = var_43_to_fp16, b = var_1395_cast_fp16, cond = mask_11)[name = string("input_249_cast_fp16")];
+            bool x_117_transpose_x_0 = const()[name = string("x_117_transpose_x_0"), val = bool(false)];
+            bool x_117_transpose_y_0 = const()[name = string("x_117_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_9_cast_fp16)[name = string("transpose_323")];
+            tensor<fp16, [1, 8, 56, 128]> x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249_cast_fp16, y = value_17_cast_fp16)[name = string("x_117_cast_fp16")];
+            tensor<int32, [4]> var_1399_perm_0 = const()[name = string("op_1399_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1400 = const()[name = string("op_1400"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_1399_cast_fp16 = transpose(perm = var_1399_perm_0, x = x_117_cast_fp16)[name = string("transpose_322")];
+            tensor<fp16, [1, 56, 1024]> input_251_cast_fp16 = reshape(shape = var_1400, x = var_1399_cast_fp16)[name = string("input_251_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103544448))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104330944))))[name = string("encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104331136)))];
+            tensor<fp16, [1, 56, 1024]> linear_43_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = string("linear_43_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = string("input_255_cast_fp16")];
+            tensor<int32, [1]> x_121_axes_0 = const()[name = string("x_121_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104333248)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104335360)))];
+            tensor<fp16, [1, 56, 1024]> x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = string("x_121_cast_fp16")];
+            tensor<int32, [3]> input_257_perm_0 = const()[name = string("input_257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_259_pad_type_0 = const()[name = string("input_259_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_259_strides_0 = const()[name = string("input_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_259_pad_0 = const()[name = string("input_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_259_dilations_0 = const()[name = string("input_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_259_groups_0 = const()[name = string("input_259_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104337472))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106434688))))[name = string("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = string("transpose_321")];
+            tensor<fp16, [1, 2048, 56]> input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("input_259_cast_fp16")];
+            int32 x_123_split_num_splits_0 = const()[name = string("x_123_split_num_splits_0"), val = int32(2)];
+            int32 x_123_split_axis_0 = const()[name = string("x_123_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_123_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_123_split_cast_fp16_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259_cast_fp16)[name = string("x_123_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = string("x_123_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = string("x_123_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_261_cast_fp16 = select(a = var_43_to_fp16, b = x_123_cast_fp16, cond = var_574)[name = string("input_261_cast_fp16")];
+            bool new_x_19_interleave_0 = const()[name = string("new_x_19_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_19_cast_fp16 = concat(axis = var_58, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = string("new_x_19_cast_fp16")];
+            tensor<int32, [3]> var_1439_begin_0 = const()[name = string("op_1439_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_1439_end_0 = const()[name = string("op_1439_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_1439_end_mask_0 = const()[name = string("op_1439_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1439_cast_fp16 = slice_by_index(begin = var_1439_begin_0, end = var_1439_end_0, end_mask = var_1439_end_mask_0, x = new_x_19_cast_fp16)[name = string("op_1439_cast_fp16")];
+            string x_125_pad_type_0 = const()[name = string("x_125_pad_type_0"), val = string("valid")];
+            int32 x_125_groups_0 = const()[name = string("x_125_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_125_strides_0 = const()[name = string("x_125_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_125_pad_0 = const()[name = string("x_125_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_125_dilations_0 = const()[name = string("x_125_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106438848))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106448128))))[name = string("encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_19_cast_fp16)[name = string("x_125_cast_fp16")];
+            tensor<int32, [3]> input_263_perm_0 = const()[name = string("input_263_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_127_axes_0 = const()[name = string("x_127_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106450240)))];
+            tensor<fp16, [1024]> encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106452352)))];
+            tensor<fp16, [1, 56, 1024]> input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = string("transpose_320")];
+            tensor<fp16, [1, 56, 1024]> x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = string("x_127_cast_fp16")];
+            tensor<int32, [3]> input_265_perm_0 = const()[name = string("input_265_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = string("transpose_319")];
+            tensor<fp16, [1, 1024, 56]> input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = string("input_267_cast_fp16")];
+            string x_129_pad_type_0 = const()[name = string("x_129_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_129_strides_0 = const()[name = string("x_129_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_129_pad_0 = const()[name = string("x_129_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_129_dilations_0 = const()[name = string("x_129_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_129_groups_0 = const()[name = string("x_129_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106454464))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107503104))))[name = string("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_129_cast_fp16 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("x_129_cast_fp16")];
+            tensor<int32, [3]> input_269_perm_0 = const()[name = string("input_269_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = string("transpose_318")];
+            tensor<fp16, [1, 56, 1024]> input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = string("input_271_cast_fp16")];
+            tensor<int32, [1]> input_273_axes_0 = const()[name = string("input_273_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107505216)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107507328)))];
+            tensor<fp16, [1, 56, 1024]> input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = string("input_273_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107509440))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110655232))))[name = string("encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_4_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110655424)))];
+            tensor<fp16, [1, 56, 4096]> linear_44_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized, x = input_273_cast_fp16)[name = string("linear_44_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = string("input_277_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110663680))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113809472))))[name = string("encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113809664)))];
+            tensor<fp16, [1, 56, 1024]> linear_45_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized, x = input_277_cast_fp16)[name = string("linear_45_cast_fp16")];
+            fp16 var_1482_to_fp16 = const()[name = string("op_1482_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1483_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1482_to_fp16)[name = string("op_1483_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1483_cast_fp16)[name = string("input_283_cast_fp16")];
+            tensor<int32, [1]> input_285_axes_0 = const()[name = string("input_285_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113811776)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113813888)))];
+            tensor<fp16, [1, 56, 1024]> input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = string("input_285_cast_fp16")];
+            tensor<int32, [4]> cache_21_begin_0 = const()[name = string("cache_21_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_21_end_0 = const()[name = string("cache_21_end_0"), val = tensor<int32, [4]>([6, 1, 42, 1024])];
+            tensor<bool, [4]> cache_21_end_mask_0 = const()[name = string("cache_21_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_21_squeeze_mask_0 = const()[name = string("cache_21_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_21_cast_fp16 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_21_cast_fp16")];
+            tensor<int32, [4]> cache_23_begin_0 = const()[name = string("cache_23_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_23_end_0 = const()[name = string("cache_23_end_0"), val = tensor<int32, [4]>([6, 1, 1024, 8])];
+            tensor<bool, [4]> cache_23_end_mask_0 = const()[name = string("cache_23_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_23_squeeze_mask_0 = const()[name = string("cache_23_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_23_cast_fp16 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_23_cast_fp16")];
+            tensor<int32, [1]> input_287_axes_0 = const()[name = string("input_287_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113816000)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113818112)))];
+            tensor<fp16, [1, 56, 1024]> input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = string("input_287_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113820224))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116966016))))[name = string("encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_5_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116966208)))];
+            tensor<fp16, [1, 56, 4096]> linear_46_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = string("linear_46_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("input_291_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116974464))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120120256))))[name = string("encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120120448)))];
+            tensor<fp16, [1, 56, 1024]> linear_47_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized, x = input_291_cast_fp16)[name = string("linear_47_cast_fp16")];
+            fp16 var_1519_to_fp16 = const()[name = string("op_1519_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1520_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1519_to_fp16)[name = string("op_1520_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1520_cast_fp16)[name = string("input_297_cast_fp16")];
+            tensor<int32, [1]> key_11_axes_0 = const()[name = string("key_11_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120122560)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120124672)))];
+            tensor<fp16, [1, 56, 1024]> key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = string("key_11_cast_fp16")];
+            bool input_299_interleave_0 = const()[name = string("input_299_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_299_cast_fp16 = concat(axis = var_67, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = string("input_299_cast_fp16")];
+            bool var_1548_interleave_0 = const()[name = string("op_1548_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_1548_cast_fp16 = concat(axis = var_67, interleave = var_1548_interleave_0, values = key_11_cast_fp16)[name = string("op_1548_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120126784))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120913280))))[name = string("encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120913472)))];
+            tensor<fp16, [1, 56, 1024]> linear_48_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized, x = key_11_cast_fp16)[name = string("linear_48_cast_fp16")];
+            tensor<int32, [4]> var_1553 = const()[name = string("op_1553"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_31_cast_fp16 = reshape(shape = var_1553, x = linear_48_cast_fp16)[name = string("q_31_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120915584))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121702080))))[name = string("encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121702272)))];
+            tensor<fp16, [1, 98, 1024]> linear_49_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_49_cast_fp16")];
+            tensor<int32, [4]> var_1558 = const()[name = string("op_1558"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_21_cast_fp16 = reshape(shape = var_1558, x = linear_49_cast_fp16)[name = string("k_21_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121704384))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122490880))))[name = string("encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122491072)))];
+            tensor<fp16, [1, 98, 1024]> linear_50_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_50_cast_fp16")];
+            tensor<int32, [4]> var_1563 = const()[name = string("op_1563"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_11_cast_fp16 = reshape(shape = var_1563, x = linear_50_cast_fp16)[name = string("v_11_cast_fp16")];
+            tensor<int32, [4]> value_19_perm_0 = const()[name = string("value_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122493184)))];
+            tensor<fp16, [1, 56, 8, 128]> var_1576_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = string("op_1576_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122495296)))];
+            tensor<fp16, [1, 56, 8, 128]> var_1578_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = string("op_1578_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_11_perm_0 = const()[name = string("q_with_bias_v_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_137_transpose_x_0 = const()[name = string("x_137_transpose_x_0"), val = bool(false)];
+            bool x_137_transpose_y_0 = const()[name = string("x_137_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_1580_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122497408))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122697152))))[name = string("op_1580_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1578_cast_fp16)[name = string("transpose_317")];
+            tensor<fp16, [1, 8, 56, 195]> x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = op_1580_to_fp16_quantized)[name = string("x_137_cast_fp16")];
+            tensor<int32, [8]> x_139_pad_0 = const()[name = string("x_139_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_139_mode_0 = const()[name = string("x_139_mode_0"), val = string("constant")];
+            fp16 const_144_to_fp16 = const()[name = string("const_144_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_139_cast_fp16 = pad(constant_val = const_144_to_fp16, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137_cast_fp16)[name = string("x_139_cast_fp16")];
+            tensor<int32, [4]> var_1588 = const()[name = string("op_1588"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_141_cast_fp16 = reshape(shape = var_1588, x = x_139_cast_fp16)[name = string("x_141_cast_fp16")];
+            tensor<int32, [4]> var_1592_begin_0 = const()[name = string("op_1592_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1592_end_0 = const()[name = string("op_1592_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_1592_end_mask_0 = const()[name = string("op_1592_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_1592_cast_fp16 = slice_by_index(begin = var_1592_begin_0, end = var_1592_end_0, end_mask = var_1592_end_mask_0, x = x_141_cast_fp16)[name = string("op_1592_cast_fp16")];
+            tensor<int32, [4]> var_1593 = const()[name = string("op_1593"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_21_cast_fp16 = reshape(shape = var_1593, x = var_1592_cast_fp16)[name = string("matrix_bd_21_cast_fp16")];
+            bool matrix_ac_11_transpose_x_0 = const()[name = string("matrix_ac_11_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_11_transpose_y_0 = const()[name = string("matrix_ac_11_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_107 = transpose(perm = transpose_107_perm_0, x = k_21_cast_fp16)[name = string("transpose_315")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_106 = transpose(perm = transpose_106_perm_0, x = var_1576_cast_fp16)[name = string("transpose_316")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_106, y = transpose_107)[name = string("matrix_ac_11_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_23_begin_0 = const()[name = string("matrix_bd_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_23_end_0 = const()[name = string("matrix_bd_23_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_23_end_mask_0 = const()[name = string("matrix_bd_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21_cast_fp16)[name = string("matrix_bd_23_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_1602_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = string("op_1602_cast_fp16")];
+            fp16 _inversed_scores_21_y_0_to_fp16 = const()[name = string("_inversed_scores_21_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_21_cast_fp16 = mul(x = var_1602_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = string("_inversed_scores_21_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_23_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_11)[name = string("scores_23_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_1608_cast_fp16 = softmax(axis = var_58, x = scores_23_cast_fp16)[name = string("op_1608_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_301_cast_fp16 = select(a = var_43_to_fp16, b = var_1608_cast_fp16, cond = mask_11)[name = string("input_301_cast_fp16")];
+            bool x_143_transpose_x_0 = const()[name = string("x_143_transpose_x_0"), val = bool(false)];
+            bool x_143_transpose_y_0 = const()[name = string("x_143_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_11_cast_fp16)[name = string("transpose_314")];
+            tensor<fp16, [1, 8, 56, 128]> x_143_cast_fp16 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301_cast_fp16, y = value_19_cast_fp16)[name = string("x_143_cast_fp16")];
+            tensor<int32, [4]> var_1612_perm_0 = const()[name = string("op_1612_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1613 = const()[name = string("op_1613"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_1612_cast_fp16 = transpose(perm = var_1612_perm_0, x = x_143_cast_fp16)[name = string("transpose_313")];
+            tensor<fp16, [1, 56, 1024]> input_303_cast_fp16 = reshape(shape = var_1613, x = var_1612_cast_fp16)[name = string("input_303_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122697664))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123484160))))[name = string("encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123484352)))];
+            tensor<fp16, [1, 56, 1024]> linear_52_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized, x = input_303_cast_fp16)[name = string("linear_52_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = string("input_307_cast_fp16")];
+            tensor<int32, [1]> x_147_axes_0 = const()[name = string("x_147_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123486464)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123488576)))];
+            tensor<fp16, [1, 56, 1024]> x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = string("x_147_cast_fp16")];
+            tensor<int32, [3]> input_309_perm_0 = const()[name = string("input_309_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_311_pad_type_0 = const()[name = string("input_311_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_311_strides_0 = const()[name = string("input_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_311_pad_0 = const()[name = string("input_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_311_dilations_0 = const()[name = string("input_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_311_groups_0 = const()[name = string("input_311_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123490688))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125587904))))[name = string("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = string("transpose_312")];
+            tensor<fp16, [1, 2048, 56]> input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_309_cast_fp16)[name = string("input_311_cast_fp16")];
+            int32 x_149_split_num_splits_0 = const()[name = string("x_149_split_num_splits_0"), val = int32(2)];
+            int32 x_149_split_axis_0 = const()[name = string("x_149_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_149_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_149_split_cast_fp16_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311_cast_fp16)[name = string("x_149_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = string("x_149_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = string("x_149_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_313_cast_fp16 = select(a = var_43_to_fp16, b = x_149_cast_fp16, cond = var_574)[name = string("input_313_cast_fp16")];
+            bool new_x_23_interleave_0 = const()[name = string("new_x_23_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_23_cast_fp16 = concat(axis = var_58, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = string("new_x_23_cast_fp16")];
+            tensor<int32, [3]> var_1652_begin_0 = const()[name = string("op_1652_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_1652_end_0 = const()[name = string("op_1652_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_1652_end_mask_0 = const()[name = string("op_1652_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1652_cast_fp16 = slice_by_index(begin = var_1652_begin_0, end = var_1652_end_0, end_mask = var_1652_end_mask_0, x = new_x_23_cast_fp16)[name = string("op_1652_cast_fp16")];
+            string x_151_pad_type_0 = const()[name = string("x_151_pad_type_0"), val = string("valid")];
+            int32 x_151_groups_0 = const()[name = string("x_151_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_151_strides_0 = const()[name = string("x_151_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_151_pad_0 = const()[name = string("x_151_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_151_dilations_0 = const()[name = string("x_151_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125592064))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125601344))))[name = string("encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_151_cast_fp16 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_23_cast_fp16)[name = string("x_151_cast_fp16")];
+            tensor<int32, [3]> input_315_perm_0 = const()[name = string("input_315_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_153_axes_0 = const()[name = string("x_153_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125603456)))];
+            tensor<fp16, [1024]> encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125605568)))];
+            tensor<fp16, [1, 56, 1024]> input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = string("transpose_311")];
+            tensor<fp16, [1, 56, 1024]> x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = string("x_153_cast_fp16")];
+            tensor<int32, [3]> input_317_perm_0 = const()[name = string("input_317_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = string("transpose_310")];
+            tensor<fp16, [1, 1024, 56]> input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = string("input_319_cast_fp16")];
+            string x_155_pad_type_0 = const()[name = string("x_155_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_155_strides_0 = const()[name = string("x_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_155_pad_0 = const()[name = string("x_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_155_dilations_0 = const()[name = string("x_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_155_groups_0 = const()[name = string("x_155_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125607680))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126656320))))[name = string("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_319_cast_fp16)[name = string("x_155_cast_fp16")];
+            tensor<int32, [3]> input_321_perm_0 = const()[name = string("input_321_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = string("transpose_309")];
+            tensor<fp16, [1, 56, 1024]> input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = string("input_323_cast_fp16")];
+            tensor<int32, [1]> input_325_axes_0 = const()[name = string("input_325_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126658432)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126660544)))];
+            tensor<fp16, [1, 56, 1024]> input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = string("input_325_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126662656))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129808448))))[name = string("encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_5_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129808640)))];
+            tensor<fp16, [1, 56, 4096]> linear_53_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized, x = input_325_cast_fp16)[name = string("linear_53_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("input_329_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129816896))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132962688))))[name = string("encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132962880)))];
+            tensor<fp16, [1, 56, 1024]> linear_54_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized, x = input_329_cast_fp16)[name = string("linear_54_cast_fp16")];
+            fp16 var_1695_to_fp16 = const()[name = string("op_1695_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1696_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1695_to_fp16)[name = string("op_1696_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1696_cast_fp16)[name = string("input_335_cast_fp16")];
+            tensor<int32, [1]> input_337_axes_0 = const()[name = string("input_337_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132964992)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132967104)))];
+            tensor<fp16, [1, 56, 1024]> input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = string("input_337_cast_fp16")];
+            tensor<int32, [4]> cache_25_begin_0 = const()[name = string("cache_25_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_25_end_0 = const()[name = string("cache_25_end_0"), val = tensor<int32, [4]>([7, 1, 42, 1024])];
+            tensor<bool, [4]> cache_25_end_mask_0 = const()[name = string("cache_25_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_25_squeeze_mask_0 = const()[name = string("cache_25_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_25_cast_fp16 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_25_cast_fp16")];
+            tensor<int32, [4]> cache_27_begin_0 = const()[name = string("cache_27_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_27_end_0 = const()[name = string("cache_27_end_0"), val = tensor<int32, [4]>([7, 1, 1024, 8])];
+            tensor<bool, [4]> cache_27_end_mask_0 = const()[name = string("cache_27_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_27_squeeze_mask_0 = const()[name = string("cache_27_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_27_cast_fp16 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_27_cast_fp16")];
+            tensor<int32, [1]> input_339_axes_0 = const()[name = string("input_339_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132969216)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132971328)))];
+            tensor<fp16, [1, 56, 1024]> input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = string("input_339_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132973440))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136119232))))[name = string("encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_6_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136119424)))];
+            tensor<fp16, [1, 56, 4096]> linear_55_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized, x = input_339_cast_fp16)[name = string("linear_55_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = string("input_343_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(136127680))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139273472))))[name = string("encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139273664)))];
+            tensor<fp16, [1, 56, 1024]> linear_56_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized, x = input_343_cast_fp16)[name = string("linear_56_cast_fp16")];
+            fp16 var_1732_to_fp16 = const()[name = string("op_1732_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1733_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1732_to_fp16)[name = string("op_1733_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1733_cast_fp16)[name = string("input_349_cast_fp16")];
+            tensor<int32, [1]> key_13_axes_0 = const()[name = string("key_13_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139275776)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139277888)))];
+            tensor<fp16, [1, 56, 1024]> key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = string("key_13_cast_fp16")];
+            bool input_351_interleave_0 = const()[name = string("input_351_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_351_cast_fp16 = concat(axis = var_67, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = string("input_351_cast_fp16")];
+            bool var_1761_interleave_0 = const()[name = string("op_1761_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_1761_cast_fp16 = concat(axis = var_67, interleave = var_1761_interleave_0, values = key_13_cast_fp16)[name = string("op_1761_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139280000))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140066496))))[name = string("encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140066688)))];
+            tensor<fp16, [1, 56, 1024]> linear_57_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized, x = key_13_cast_fp16)[name = string("linear_57_cast_fp16")];
+            tensor<int32, [4]> var_1766 = const()[name = string("op_1766"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_37_cast_fp16 = reshape(shape = var_1766, x = linear_57_cast_fp16)[name = string("q_37_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140068800))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140855296))))[name = string("encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140855488)))];
+            tensor<fp16, [1, 98, 1024]> linear_58_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_58_cast_fp16")];
+            tensor<int32, [4]> var_1771 = const()[name = string("op_1771"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_25_cast_fp16 = reshape(shape = var_1771, x = linear_58_cast_fp16)[name = string("k_25_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140857600))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141644096))))[name = string("encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141644288)))];
+            tensor<fp16, [1, 98, 1024]> linear_59_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_59_cast_fp16")];
+            tensor<int32, [4]> var_1776 = const()[name = string("op_1776"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_13_cast_fp16 = reshape(shape = var_1776, x = linear_59_cast_fp16)[name = string("v_13_cast_fp16")];
+            tensor<int32, [4]> value_21_perm_0 = const()[name = string("value_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141646400)))];
+            tensor<fp16, [1, 56, 8, 128]> var_1789_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = string("op_1789_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141648512)))];
+            tensor<fp16, [1, 56, 8, 128]> var_1791_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = string("op_1791_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_13_perm_0 = const()[name = string("q_with_bias_v_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_163_transpose_x_0 = const()[name = string("x_163_transpose_x_0"), val = bool(false)];
+            bool x_163_transpose_y_0 = const()[name = string("x_163_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_1793_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141650624))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141850368))))[name = string("op_1793_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1791_cast_fp16)[name = string("transpose_308")];
+            tensor<fp16, [1, 8, 56, 195]> x_163_cast_fp16 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = op_1793_to_fp16_quantized)[name = string("x_163_cast_fp16")];
+            tensor<int32, [8]> x_165_pad_0 = const()[name = string("x_165_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_165_mode_0 = const()[name = string("x_165_mode_0"), val = string("constant")];
+            fp16 const_157_to_fp16 = const()[name = string("const_157_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_165_cast_fp16 = pad(constant_val = const_157_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = string("x_165_cast_fp16")];
+            tensor<int32, [4]> var_1801 = const()[name = string("op_1801"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_167_cast_fp16 = reshape(shape = var_1801, x = x_165_cast_fp16)[name = string("x_167_cast_fp16")];
+            tensor<int32, [4]> var_1805_begin_0 = const()[name = string("op_1805_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1805_end_0 = const()[name = string("op_1805_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_1805_end_mask_0 = const()[name = string("op_1805_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_1805_cast_fp16 = slice_by_index(begin = var_1805_begin_0, end = var_1805_end_0, end_mask = var_1805_end_mask_0, x = x_167_cast_fp16)[name = string("op_1805_cast_fp16")];
+            tensor<int32, [4]> var_1806 = const()[name = string("op_1806"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_25_cast_fp16 = reshape(shape = var_1806, x = var_1805_cast_fp16)[name = string("matrix_bd_25_cast_fp16")];
+            bool matrix_ac_13_transpose_x_0 = const()[name = string("matrix_ac_13_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_13_transpose_y_0 = const()[name = string("matrix_ac_13_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_109 = transpose(perm = transpose_109_perm_0, x = k_25_cast_fp16)[name = string("transpose_306")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1789_cast_fp16)[name = string("transpose_307")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_108, y = transpose_109)[name = string("matrix_ac_13_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_27_begin_0 = const()[name = string("matrix_bd_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_27_end_0 = const()[name = string("matrix_bd_27_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_27_end_mask_0 = const()[name = string("matrix_bd_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25_cast_fp16)[name = string("matrix_bd_27_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_1815_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = string("op_1815_cast_fp16")];
+            fp16 _inversed_scores_25_y_0_to_fp16 = const()[name = string("_inversed_scores_25_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_25_cast_fp16 = mul(x = var_1815_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = string("_inversed_scores_25_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_27_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_11)[name = string("scores_27_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_1821_cast_fp16 = softmax(axis = var_58, x = scores_27_cast_fp16)[name = string("op_1821_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_353_cast_fp16 = select(a = var_43_to_fp16, b = var_1821_cast_fp16, cond = mask_11)[name = string("input_353_cast_fp16")];
+            bool x_169_transpose_x_0 = const()[name = string("x_169_transpose_x_0"), val = bool(false)];
+            bool x_169_transpose_y_0 = const()[name = string("x_169_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_13_cast_fp16)[name = string("transpose_305")];
+            tensor<fp16, [1, 8, 56, 128]> x_169_cast_fp16 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353_cast_fp16, y = value_21_cast_fp16)[name = string("x_169_cast_fp16")];
+            tensor<int32, [4]> var_1825_perm_0 = const()[name = string("op_1825_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1826 = const()[name = string("op_1826"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_1825_cast_fp16 = transpose(perm = var_1825_perm_0, x = x_169_cast_fp16)[name = string("transpose_304")];
+            tensor<fp16, [1, 56, 1024]> input_355_cast_fp16 = reshape(shape = var_1826, x = var_1825_cast_fp16)[name = string("input_355_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141850880))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142637376))))[name = string("encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142637568)))];
+            tensor<fp16, [1, 56, 1024]> linear_61_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized, x = input_355_cast_fp16)[name = string("linear_61_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = string("input_359_cast_fp16")];
+            tensor<int32, [1]> x_173_axes_0 = const()[name = string("x_173_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142639680)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142641792)))];
+            tensor<fp16, [1, 56, 1024]> x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = string("x_173_cast_fp16")];
+            tensor<int32, [3]> input_361_perm_0 = const()[name = string("input_361_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_363_pad_type_0 = const()[name = string("input_363_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_363_strides_0 = const()[name = string("input_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_363_pad_0 = const()[name = string("input_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_363_dilations_0 = const()[name = string("input_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_363_groups_0 = const()[name = string("input_363_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142643904))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144741120))))[name = string("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = string("transpose_303")];
+            tensor<fp16, [1, 2048, 56]> input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_361_cast_fp16)[name = string("input_363_cast_fp16")];
+            int32 x_175_split_num_splits_0 = const()[name = string("x_175_split_num_splits_0"), val = int32(2)];
+            int32 x_175_split_axis_0 = const()[name = string("x_175_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_175_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_175_split_cast_fp16_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363_cast_fp16)[name = string("x_175_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = string("x_175_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = string("x_175_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_365_cast_fp16 = select(a = var_43_to_fp16, b = x_175_cast_fp16, cond = var_574)[name = string("input_365_cast_fp16")];
+            bool new_x_27_interleave_0 = const()[name = string("new_x_27_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_27_cast_fp16 = concat(axis = var_58, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = string("new_x_27_cast_fp16")];
+            tensor<int32, [3]> var_1865_begin_0 = const()[name = string("op_1865_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_1865_end_0 = const()[name = string("op_1865_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_1865_end_mask_0 = const()[name = string("op_1865_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1865_cast_fp16 = slice_by_index(begin = var_1865_begin_0, end = var_1865_end_0, end_mask = var_1865_end_mask_0, x = new_x_27_cast_fp16)[name = string("op_1865_cast_fp16")];
+            string x_177_pad_type_0 = const()[name = string("x_177_pad_type_0"), val = string("valid")];
+            int32 x_177_groups_0 = const()[name = string("x_177_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_177_strides_0 = const()[name = string("x_177_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_177_pad_0 = const()[name = string("x_177_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_177_dilations_0 = const()[name = string("x_177_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144745280))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144754560))))[name = string("encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_177_cast_fp16 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_27_cast_fp16)[name = string("x_177_cast_fp16")];
+            tensor<int32, [3]> input_367_perm_0 = const()[name = string("input_367_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_179_axes_0 = const()[name = string("x_179_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144756672)))];
+            tensor<fp16, [1024]> encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144758784)))];
+            tensor<fp16, [1, 56, 1024]> input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = string("transpose_302")];
+            tensor<fp16, [1, 56, 1024]> x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = string("x_179_cast_fp16")];
+            tensor<int32, [3]> input_369_perm_0 = const()[name = string("input_369_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = string("transpose_301")];
+            tensor<fp16, [1, 1024, 56]> input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = string("input_371_cast_fp16")];
+            string x_181_pad_type_0 = const()[name = string("x_181_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_181_strides_0 = const()[name = string("x_181_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_181_pad_0 = const()[name = string("x_181_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_181_dilations_0 = const()[name = string("x_181_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_181_groups_0 = const()[name = string("x_181_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144760896))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145809536))))[name = string("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_181_cast_fp16 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_371_cast_fp16)[name = string("x_181_cast_fp16")];
+            tensor<int32, [3]> input_373_perm_0 = const()[name = string("input_373_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = string("transpose_300")];
+            tensor<fp16, [1, 56, 1024]> input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = string("input_375_cast_fp16")];
+            tensor<int32, [1]> input_377_axes_0 = const()[name = string("input_377_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145811648)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145813760)))];
+            tensor<fp16, [1, 56, 1024]> input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = string("input_377_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145815872))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148961664))))[name = string("encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_6_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148961856)))];
+            tensor<fp16, [1, 56, 4096]> linear_62_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized, x = input_377_cast_fp16)[name = string("linear_62_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = string("input_381_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148970112))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152115904))))[name = string("encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152116096)))];
+            tensor<fp16, [1, 56, 1024]> linear_63_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized, x = input_381_cast_fp16)[name = string("linear_63_cast_fp16")];
+            fp16 var_1908_to_fp16 = const()[name = string("op_1908_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1909_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1908_to_fp16)[name = string("op_1909_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1909_cast_fp16)[name = string("input_387_cast_fp16")];
+            tensor<int32, [1]> input_389_axes_0 = const()[name = string("input_389_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152118208)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152120320)))];
+            tensor<fp16, [1, 56, 1024]> input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = string("input_389_cast_fp16")];
+            tensor<int32, [4]> cache_29_begin_0 = const()[name = string("cache_29_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_29_end_0 = const()[name = string("cache_29_end_0"), val = tensor<int32, [4]>([8, 1, 42, 1024])];
+            tensor<bool, [4]> cache_29_end_mask_0 = const()[name = string("cache_29_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_29_squeeze_mask_0 = const()[name = string("cache_29_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_29_cast_fp16 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_29_cast_fp16")];
+            tensor<int32, [4]> cache_31_begin_0 = const()[name = string("cache_31_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_31_end_0 = const()[name = string("cache_31_end_0"), val = tensor<int32, [4]>([8, 1, 1024, 8])];
+            tensor<bool, [4]> cache_31_end_mask_0 = const()[name = string("cache_31_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_31_squeeze_mask_0 = const()[name = string("cache_31_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_31_cast_fp16 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_31_cast_fp16")];
+            tensor<int32, [1]> input_391_axes_0 = const()[name = string("input_391_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152122432)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152124544)))];
+            tensor<fp16, [1, 56, 1024]> input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = string("input_391_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152126656))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(155272448))))[name = string("encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_7_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(155272640)))];
+            tensor<fp16, [1, 56, 4096]> linear_64_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized, x = input_391_cast_fp16)[name = string("linear_64_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = string("input_395_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(155280896))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158426688))))[name = string("encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158426880)))];
+            tensor<fp16, [1, 56, 1024]> linear_65_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized, x = input_395_cast_fp16)[name = string("linear_65_cast_fp16")];
+            fp16 var_1945_to_fp16 = const()[name = string("op_1945_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_1946_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1945_to_fp16)[name = string("op_1946_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1946_cast_fp16)[name = string("input_401_cast_fp16")];
+            tensor<int32, [1]> key_15_axes_0 = const()[name = string("key_15_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158428992)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158431104)))];
+            tensor<fp16, [1, 56, 1024]> key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = string("key_15_cast_fp16")];
+            bool input_403_interleave_0 = const()[name = string("input_403_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_403_cast_fp16 = concat(axis = var_67, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = string("input_403_cast_fp16")];
+            bool var_1974_interleave_0 = const()[name = string("op_1974_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_1974_cast_fp16 = concat(axis = var_67, interleave = var_1974_interleave_0, values = key_15_cast_fp16)[name = string("op_1974_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158433216))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159219712))))[name = string("encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159219904)))];
+            tensor<fp16, [1, 56, 1024]> linear_66_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized, x = key_15_cast_fp16)[name = string("linear_66_cast_fp16")];
+            tensor<int32, [4]> var_1979 = const()[name = string("op_1979"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_43_cast_fp16 = reshape(shape = var_1979, x = linear_66_cast_fp16)[name = string("q_43_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159222016))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160008512))))[name = string("encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160008704)))];
+            tensor<fp16, [1, 98, 1024]> linear_67_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_67_cast_fp16")];
+            tensor<int32, [4]> var_1984 = const()[name = string("op_1984"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_29_cast_fp16 = reshape(shape = var_1984, x = linear_67_cast_fp16)[name = string("k_29_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160010816))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160797312))))[name = string("encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160797504)))];
+            tensor<fp16, [1, 98, 1024]> linear_68_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_68_cast_fp16")];
+            tensor<int32, [4]> var_1989 = const()[name = string("op_1989"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_15_cast_fp16 = reshape(shape = var_1989, x = linear_68_cast_fp16)[name = string("v_15_cast_fp16")];
+            tensor<int32, [4]> value_23_perm_0 = const()[name = string("value_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160799616)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2002_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = string("op_2002_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160801728)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2004_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = string("op_2004_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_15_perm_0 = const()[name = string("q_with_bias_v_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_189_transpose_x_0 = const()[name = string("x_189_transpose_x_0"), val = bool(false)];
+            bool x_189_transpose_y_0 = const()[name = string("x_189_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_2006_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160803840))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161003584))))[name = string("op_2006_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_2004_cast_fp16)[name = string("transpose_299")];
+            tensor<fp16, [1, 8, 56, 195]> x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = op_2006_to_fp16_quantized)[name = string("x_189_cast_fp16")];
+            tensor<int32, [8]> x_191_pad_0 = const()[name = string("x_191_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_191_mode_0 = const()[name = string("x_191_mode_0"), val = string("constant")];
+            fp16 const_170_to_fp16 = const()[name = string("const_170_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_191_cast_fp16 = pad(constant_val = const_170_to_fp16, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189_cast_fp16)[name = string("x_191_cast_fp16")];
+            tensor<int32, [4]> var_2014 = const()[name = string("op_2014"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_193_cast_fp16 = reshape(shape = var_2014, x = x_191_cast_fp16)[name = string("x_193_cast_fp16")];
+            tensor<int32, [4]> var_2018_begin_0 = const()[name = string("op_2018_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2018_end_0 = const()[name = string("op_2018_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_2018_end_mask_0 = const()[name = string("op_2018_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_2018_cast_fp16 = slice_by_index(begin = var_2018_begin_0, end = var_2018_end_0, end_mask = var_2018_end_mask_0, x = x_193_cast_fp16)[name = string("op_2018_cast_fp16")];
+            tensor<int32, [4]> var_2019 = const()[name = string("op_2019"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_29_cast_fp16 = reshape(shape = var_2019, x = var_2018_cast_fp16)[name = string("matrix_bd_29_cast_fp16")];
+            bool matrix_ac_15_transpose_x_0 = const()[name = string("matrix_ac_15_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_15_transpose_y_0 = const()[name = string("matrix_ac_15_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_111 = transpose(perm = transpose_111_perm_0, x = k_29_cast_fp16)[name = string("transpose_297")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_110 = transpose(perm = transpose_110_perm_0, x = var_2002_cast_fp16)[name = string("transpose_298")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_110, y = transpose_111)[name = string("matrix_ac_15_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_31_begin_0 = const()[name = string("matrix_bd_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_31_end_0 = const()[name = string("matrix_bd_31_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_31_end_mask_0 = const()[name = string("matrix_bd_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29_cast_fp16)[name = string("matrix_bd_31_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2028_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = string("op_2028_cast_fp16")];
+            fp16 _inversed_scores_29_y_0_to_fp16 = const()[name = string("_inversed_scores_29_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_29_cast_fp16 = mul(x = var_2028_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = string("_inversed_scores_29_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_31_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_11)[name = string("scores_31_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2034_cast_fp16 = softmax(axis = var_58, x = scores_31_cast_fp16)[name = string("op_2034_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_405_cast_fp16 = select(a = var_43_to_fp16, b = var_2034_cast_fp16, cond = mask_11)[name = string("input_405_cast_fp16")];
+            bool x_195_transpose_x_0 = const()[name = string("x_195_transpose_x_0"), val = bool(false)];
+            bool x_195_transpose_y_0 = const()[name = string("x_195_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_15_cast_fp16)[name = string("transpose_296")];
+            tensor<fp16, [1, 8, 56, 128]> x_195_cast_fp16 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405_cast_fp16, y = value_23_cast_fp16)[name = string("x_195_cast_fp16")];
+            tensor<int32, [4]> var_2038_perm_0 = const()[name = string("op_2038_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2039 = const()[name = string("op_2039"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_2038_cast_fp16 = transpose(perm = var_2038_perm_0, x = x_195_cast_fp16)[name = string("transpose_295")];
+            tensor<fp16, [1, 56, 1024]> input_407_cast_fp16 = reshape(shape = var_2039, x = var_2038_cast_fp16)[name = string("input_407_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161004096))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161790592))))[name = string("encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161790784)))];
+            tensor<fp16, [1, 56, 1024]> linear_70_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = string("linear_70_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = string("input_411_cast_fp16")];
+            tensor<int32, [1]> x_199_axes_0 = const()[name = string("x_199_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161792896)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161795008)))];
+            tensor<fp16, [1, 56, 1024]> x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = string("x_199_cast_fp16")];
+            tensor<int32, [3]> input_413_perm_0 = const()[name = string("input_413_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_415_pad_type_0 = const()[name = string("input_415_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_415_strides_0 = const()[name = string("input_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_415_pad_0 = const()[name = string("input_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_415_dilations_0 = const()[name = string("input_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_415_groups_0 = const()[name = string("input_415_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161797120))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163894336))))[name = string("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = string("transpose_294")];
+            tensor<fp16, [1, 2048, 56]> input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_413_cast_fp16)[name = string("input_415_cast_fp16")];
+            int32 x_201_split_num_splits_0 = const()[name = string("x_201_split_num_splits_0"), val = int32(2)];
+            int32 x_201_split_axis_0 = const()[name = string("x_201_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_201_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_201_split_cast_fp16_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415_cast_fp16)[name = string("x_201_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = string("x_201_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = string("x_201_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_417_cast_fp16 = select(a = var_43_to_fp16, b = x_201_cast_fp16, cond = var_574)[name = string("input_417_cast_fp16")];
+            bool new_x_31_interleave_0 = const()[name = string("new_x_31_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_31_cast_fp16 = concat(axis = var_58, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = string("new_x_31_cast_fp16")];
+            tensor<int32, [3]> var_2078_begin_0 = const()[name = string("op_2078_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_2078_end_0 = const()[name = string("op_2078_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_2078_end_mask_0 = const()[name = string("op_2078_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2078_cast_fp16 = slice_by_index(begin = var_2078_begin_0, end = var_2078_end_0, end_mask = var_2078_end_mask_0, x = new_x_31_cast_fp16)[name = string("op_2078_cast_fp16")];
+            string x_203_pad_type_0 = const()[name = string("x_203_pad_type_0"), val = string("valid")];
+            int32 x_203_groups_0 = const()[name = string("x_203_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_203_strides_0 = const()[name = string("x_203_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_203_pad_0 = const()[name = string("x_203_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_203_dilations_0 = const()[name = string("x_203_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163898496))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163907776))))[name = string("encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_31_cast_fp16)[name = string("x_203_cast_fp16")];
+            tensor<int32, [3]> input_419_perm_0 = const()[name = string("input_419_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_205_axes_0 = const()[name = string("x_205_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163909888)))];
+            tensor<fp16, [1024]> encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163912000)))];
+            tensor<fp16, [1, 56, 1024]> input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = string("transpose_293")];
+            tensor<fp16, [1, 56, 1024]> x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = string("x_205_cast_fp16")];
+            tensor<int32, [3]> input_421_perm_0 = const()[name = string("input_421_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = string("transpose_292")];
+            tensor<fp16, [1, 1024, 56]> input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = string("input_423_cast_fp16")];
+            string x_207_pad_type_0 = const()[name = string("x_207_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_207_strides_0 = const()[name = string("x_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_207_pad_0 = const()[name = string("x_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_207_dilations_0 = const()[name = string("x_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_207_groups_0 = const()[name = string("x_207_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163914112))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164962752))))[name = string("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_207_cast_fp16 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_423_cast_fp16)[name = string("x_207_cast_fp16")];
+            tensor<int32, [3]> input_425_perm_0 = const()[name = string("input_425_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = string("transpose_291")];
+            tensor<fp16, [1, 56, 1024]> input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = string("input_427_cast_fp16")];
+            tensor<int32, [1]> input_429_axes_0 = const()[name = string("input_429_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164964864)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164966976)))];
+            tensor<fp16, [1, 56, 1024]> input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = string("input_429_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164969088))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168114880))))[name = string("encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_7_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168115072)))];
+            tensor<fp16, [1, 56, 4096]> linear_71_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized, x = input_429_cast_fp16)[name = string("linear_71_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = string("input_433_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(168123328))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171269120))))[name = string("encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171269312)))];
+            tensor<fp16, [1, 56, 1024]> linear_72_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized, x = input_433_cast_fp16)[name = string("linear_72_cast_fp16")];
+            fp16 var_2121_to_fp16 = const()[name = string("op_2121_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_2122_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_2121_to_fp16)[name = string("op_2122_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_2122_cast_fp16)[name = string("input_439_cast_fp16")];
+            tensor<int32, [1]> input_441_axes_0 = const()[name = string("input_441_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171271424)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171273536)))];
+            tensor<fp16, [1, 56, 1024]> input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = string("input_441_cast_fp16")];
+            tensor<int32, [4]> cache_33_begin_0 = const()[name = string("cache_33_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_33_end_0 = const()[name = string("cache_33_end_0"), val = tensor<int32, [4]>([9, 1, 42, 1024])];
+            tensor<bool, [4]> cache_33_end_mask_0 = const()[name = string("cache_33_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_33_squeeze_mask_0 = const()[name = string("cache_33_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_33_cast_fp16 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_33_cast_fp16")];
+            tensor<int32, [4]> cache_35_begin_0 = const()[name = string("cache_35_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_35_end_0 = const()[name = string("cache_35_end_0"), val = tensor<int32, [4]>([9, 1, 1024, 8])];
+            tensor<bool, [4]> cache_35_end_mask_0 = const()[name = string("cache_35_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_35_squeeze_mask_0 = const()[name = string("cache_35_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_35_cast_fp16 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_35_cast_fp16")];
+            tensor<int32, [1]> input_443_axes_0 = const()[name = string("input_443_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171275648)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171277760)))];
+            tensor<fp16, [1, 56, 1024]> input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = string("input_443_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171279872))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174425664))))[name = string("encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_8_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174425856)))];
+            tensor<fp16, [1, 56, 4096]> linear_73_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("linear_73_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = string("input_447_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(174434112))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177579904))))[name = string("encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177580096)))];
+            tensor<fp16, [1, 56, 1024]> linear_74_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized, x = input_447_cast_fp16)[name = string("linear_74_cast_fp16")];
+            fp16 var_2158_to_fp16 = const()[name = string("op_2158_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_2159_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_2158_to_fp16)[name = string("op_2159_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_2159_cast_fp16)[name = string("input_453_cast_fp16")];
+            tensor<int32, [1]> key_17_axes_0 = const()[name = string("key_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177582208)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177584320)))];
+            tensor<fp16, [1, 56, 1024]> key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = string("key_17_cast_fp16")];
+            bool input_455_interleave_0 = const()[name = string("input_455_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_455_cast_fp16 = concat(axis = var_67, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = string("input_455_cast_fp16")];
+            bool var_2187_interleave_0 = const()[name = string("op_2187_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_2187_cast_fp16 = concat(axis = var_67, interleave = var_2187_interleave_0, values = key_17_cast_fp16)[name = string("op_2187_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177586432))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178372928))))[name = string("encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178373120)))];
+            tensor<fp16, [1, 56, 1024]> linear_75_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized, x = key_17_cast_fp16)[name = string("linear_75_cast_fp16")];
+            tensor<int32, [4]> var_2192 = const()[name = string("op_2192"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_49_cast_fp16 = reshape(shape = var_2192, x = linear_75_cast_fp16)[name = string("q_49_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178375232))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179161728))))[name = string("encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179161920)))];
+            tensor<fp16, [1, 98, 1024]> linear_76_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_76_cast_fp16")];
+            tensor<int32, [4]> var_2197 = const()[name = string("op_2197"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_33_cast_fp16 = reshape(shape = var_2197, x = linear_76_cast_fp16)[name = string("k_33_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179164032))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179950528))))[name = string("encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179950720)))];
+            tensor<fp16, [1, 98, 1024]> linear_77_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_77_cast_fp16")];
+            tensor<int32, [4]> var_2202 = const()[name = string("op_2202"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_17_cast_fp16 = reshape(shape = var_2202, x = linear_77_cast_fp16)[name = string("v_17_cast_fp16")];
+            tensor<int32, [4]> value_25_perm_0 = const()[name = string("value_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179952832)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2215_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = string("op_2215_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179954944)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2217_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = string("op_2217_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_17_perm_0 = const()[name = string("q_with_bias_v_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_215_transpose_x_0 = const()[name = string("x_215_transpose_x_0"), val = bool(false)];
+            bool x_215_transpose_y_0 = const()[name = string("x_215_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_2219_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179957056))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180156800))))[name = string("op_2219_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2217_cast_fp16)[name = string("transpose_290")];
+            tensor<fp16, [1, 8, 56, 195]> x_215_cast_fp16 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = op_2219_to_fp16_quantized)[name = string("x_215_cast_fp16")];
+            tensor<int32, [8]> x_217_pad_0 = const()[name = string("x_217_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_217_mode_0 = const()[name = string("x_217_mode_0"), val = string("constant")];
+            fp16 const_183_to_fp16 = const()[name = string("const_183_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_217_cast_fp16 = pad(constant_val = const_183_to_fp16, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215_cast_fp16)[name = string("x_217_cast_fp16")];
+            tensor<int32, [4]> var_2227 = const()[name = string("op_2227"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_219_cast_fp16 = reshape(shape = var_2227, x = x_217_cast_fp16)[name = string("x_219_cast_fp16")];
+            tensor<int32, [4]> var_2231_begin_0 = const()[name = string("op_2231_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2231_end_0 = const()[name = string("op_2231_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_2231_end_mask_0 = const()[name = string("op_2231_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_2231_cast_fp16 = slice_by_index(begin = var_2231_begin_0, end = var_2231_end_0, end_mask = var_2231_end_mask_0, x = x_219_cast_fp16)[name = string("op_2231_cast_fp16")];
+            tensor<int32, [4]> var_2232 = const()[name = string("op_2232"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_33_cast_fp16 = reshape(shape = var_2232, x = var_2231_cast_fp16)[name = string("matrix_bd_33_cast_fp16")];
+            bool matrix_ac_17_transpose_x_0 = const()[name = string("matrix_ac_17_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_17_transpose_y_0 = const()[name = string("matrix_ac_17_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_113 = transpose(perm = transpose_113_perm_0, x = k_33_cast_fp16)[name = string("transpose_288")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_112 = transpose(perm = transpose_112_perm_0, x = var_2215_cast_fp16)[name = string("transpose_289")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_112, y = transpose_113)[name = string("matrix_ac_17_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_35_begin_0 = const()[name = string("matrix_bd_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_35_end_0 = const()[name = string("matrix_bd_35_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_35_end_mask_0 = const()[name = string("matrix_bd_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33_cast_fp16)[name = string("matrix_bd_35_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2241_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = string("op_2241_cast_fp16")];
+            fp16 _inversed_scores_33_y_0_to_fp16 = const()[name = string("_inversed_scores_33_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_33_cast_fp16 = mul(x = var_2241_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = string("_inversed_scores_33_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_35_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_11)[name = string("scores_35_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2247_cast_fp16 = softmax(axis = var_58, x = scores_35_cast_fp16)[name = string("op_2247_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_457_cast_fp16 = select(a = var_43_to_fp16, b = var_2247_cast_fp16, cond = mask_11)[name = string("input_457_cast_fp16")];
+            bool x_221_transpose_x_0 = const()[name = string("x_221_transpose_x_0"), val = bool(false)];
+            bool x_221_transpose_y_0 = const()[name = string("x_221_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_17_cast_fp16)[name = string("transpose_287")];
+            tensor<fp16, [1, 8, 56, 128]> x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457_cast_fp16, y = value_25_cast_fp16)[name = string("x_221_cast_fp16")];
+            tensor<int32, [4]> var_2251_perm_0 = const()[name = string("op_2251_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2252 = const()[name = string("op_2252"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_2251_cast_fp16 = transpose(perm = var_2251_perm_0, x = x_221_cast_fp16)[name = string("transpose_286")];
+            tensor<fp16, [1, 56, 1024]> input_459_cast_fp16 = reshape(shape = var_2252, x = var_2251_cast_fp16)[name = string("input_459_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180157312))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180943808))))[name = string("encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180944000)))];
+            tensor<fp16, [1, 56, 1024]> linear_79_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized, x = input_459_cast_fp16)[name = string("linear_79_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = string("input_463_cast_fp16")];
+            tensor<int32, [1]> x_225_axes_0 = const()[name = string("x_225_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180946112)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180948224)))];
+            tensor<fp16, [1, 56, 1024]> x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = string("x_225_cast_fp16")];
+            tensor<int32, [3]> input_465_perm_0 = const()[name = string("input_465_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_467_pad_type_0 = const()[name = string("input_467_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_467_strides_0 = const()[name = string("input_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_467_pad_0 = const()[name = string("input_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_467_dilations_0 = const()[name = string("input_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_467_groups_0 = const()[name = string("input_467_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180950336))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183047552))))[name = string("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = string("transpose_285")];
+            tensor<fp16, [1, 2048, 56]> input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_465_cast_fp16)[name = string("input_467_cast_fp16")];
+            int32 x_227_split_num_splits_0 = const()[name = string("x_227_split_num_splits_0"), val = int32(2)];
+            int32 x_227_split_axis_0 = const()[name = string("x_227_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_227_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_227_split_cast_fp16_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467_cast_fp16)[name = string("x_227_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = string("x_227_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = string("x_227_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_469_cast_fp16 = select(a = var_43_to_fp16, b = x_227_cast_fp16, cond = var_574)[name = string("input_469_cast_fp16")];
+            bool new_x_35_interleave_0 = const()[name = string("new_x_35_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_35_cast_fp16 = concat(axis = var_58, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = string("new_x_35_cast_fp16")];
+            tensor<int32, [3]> var_2291_begin_0 = const()[name = string("op_2291_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_2291_end_0 = const()[name = string("op_2291_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_2291_end_mask_0 = const()[name = string("op_2291_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2291_cast_fp16 = slice_by_index(begin = var_2291_begin_0, end = var_2291_end_0, end_mask = var_2291_end_mask_0, x = new_x_35_cast_fp16)[name = string("op_2291_cast_fp16")];
+            string x_229_pad_type_0 = const()[name = string("x_229_pad_type_0"), val = string("valid")];
+            int32 x_229_groups_0 = const()[name = string("x_229_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_229_strides_0 = const()[name = string("x_229_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_229_pad_0 = const()[name = string("x_229_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_229_dilations_0 = const()[name = string("x_229_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183051712))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183060992))))[name = string("encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_229_cast_fp16 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_35_cast_fp16)[name = string("x_229_cast_fp16")];
+            tensor<int32, [3]> input_471_perm_0 = const()[name = string("input_471_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_231_axes_0 = const()[name = string("x_231_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183063104)))];
+            tensor<fp16, [1024]> encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183065216)))];
+            tensor<fp16, [1, 56, 1024]> input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = string("transpose_284")];
+            tensor<fp16, [1, 56, 1024]> x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = string("x_231_cast_fp16")];
+            tensor<int32, [3]> input_473_perm_0 = const()[name = string("input_473_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = string("transpose_283")];
+            tensor<fp16, [1, 1024, 56]> input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = string("input_475_cast_fp16")];
+            string x_233_pad_type_0 = const()[name = string("x_233_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_233_strides_0 = const()[name = string("x_233_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_233_pad_0 = const()[name = string("x_233_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_233_dilations_0 = const()[name = string("x_233_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_233_groups_0 = const()[name = string("x_233_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183067328))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184115968))))[name = string("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_233_cast_fp16 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_475_cast_fp16)[name = string("x_233_cast_fp16")];
+            tensor<int32, [3]> input_477_perm_0 = const()[name = string("input_477_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = string("transpose_282")];
+            tensor<fp16, [1, 56, 1024]> input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = string("input_479_cast_fp16")];
+            tensor<int32, [1]> input_481_axes_0 = const()[name = string("input_481_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184118080)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184120192)))];
+            tensor<fp16, [1, 56, 1024]> input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = string("input_481_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184122304))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187268096))))[name = string("encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_8_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187268288)))];
+            tensor<fp16, [1, 56, 4096]> linear_80_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized, x = input_481_cast_fp16)[name = string("linear_80_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = string("input_485_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(187276544))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190422336))))[name = string("encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190422528)))];
+            tensor<fp16, [1, 56, 1024]> linear_81_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized, x = input_485_cast_fp16)[name = string("linear_81_cast_fp16")];
+            fp16 var_2334_to_fp16 = const()[name = string("op_2334_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_2335_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2334_to_fp16)[name = string("op_2335_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2335_cast_fp16)[name = string("input_491_cast_fp16")];
+            tensor<int32, [1]> input_493_axes_0 = const()[name = string("input_493_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190424640)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190426752)))];
+            tensor<fp16, [1, 56, 1024]> input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = string("input_493_cast_fp16")];
+            tensor<int32, [4]> cache_37_begin_0 = const()[name = string("cache_37_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_37_end_0 = const()[name = string("cache_37_end_0"), val = tensor<int32, [4]>([10, 1, 42, 1024])];
+            tensor<bool, [4]> cache_37_end_mask_0 = const()[name = string("cache_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_37_squeeze_mask_0 = const()[name = string("cache_37_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_37_cast_fp16 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_37_cast_fp16")];
+            tensor<int32, [4]> cache_39_begin_0 = const()[name = string("cache_39_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_39_end_0 = const()[name = string("cache_39_end_0"), val = tensor<int32, [4]>([10, 1, 1024, 8])];
+            tensor<bool, [4]> cache_39_end_mask_0 = const()[name = string("cache_39_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_39_squeeze_mask_0 = const()[name = string("cache_39_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_39_cast_fp16 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_39_cast_fp16")];
+            tensor<int32, [1]> input_495_axes_0 = const()[name = string("input_495_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190428864)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190430976)))];
+            tensor<fp16, [1, 56, 1024]> input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = string("input_495_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190433088))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193578880))))[name = string("encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_9_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193579072)))];
+            tensor<fp16, [1, 56, 4096]> linear_82_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized, x = input_495_cast_fp16)[name = string("linear_82_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = string("input_499_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193587328))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196733120))))[name = string("encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196733312)))];
+            tensor<fp16, [1, 56, 1024]> linear_83_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized, x = input_499_cast_fp16)[name = string("linear_83_cast_fp16")];
+            fp16 var_2371_to_fp16 = const()[name = string("op_2371_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_2372_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2371_to_fp16)[name = string("op_2372_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2372_cast_fp16)[name = string("input_505_cast_fp16")];
+            tensor<int32, [1]> key_19_axes_0 = const()[name = string("key_19_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196735424)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196737536)))];
+            tensor<fp16, [1, 56, 1024]> key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = string("key_19_cast_fp16")];
+            bool input_507_interleave_0 = const()[name = string("input_507_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_507_cast_fp16 = concat(axis = var_67, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = string("input_507_cast_fp16")];
+            bool var_2400_interleave_0 = const()[name = string("op_2400_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_2400_cast_fp16 = concat(axis = var_67, interleave = var_2400_interleave_0, values = key_19_cast_fp16)[name = string("op_2400_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196739648))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197526144))))[name = string("encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197526336)))];
+            tensor<fp16, [1, 56, 1024]> linear_84_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized, x = key_19_cast_fp16)[name = string("linear_84_cast_fp16")];
+            tensor<int32, [4]> var_2405 = const()[name = string("op_2405"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_55_cast_fp16 = reshape(shape = var_2405, x = linear_84_cast_fp16)[name = string("q_55_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197528448))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198314944))))[name = string("encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198315136)))];
+            tensor<fp16, [1, 98, 1024]> linear_85_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_85_cast_fp16")];
+            tensor<int32, [4]> var_2410 = const()[name = string("op_2410"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_37_cast_fp16 = reshape(shape = var_2410, x = linear_85_cast_fp16)[name = string("k_37_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198317248))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199103744))))[name = string("encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199103936)))];
+            tensor<fp16, [1, 98, 1024]> linear_86_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_86_cast_fp16")];
+            tensor<int32, [4]> var_2415 = const()[name = string("op_2415"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_19_cast_fp16 = reshape(shape = var_2415, x = linear_86_cast_fp16)[name = string("v_19_cast_fp16")];
+            tensor<int32, [4]> value_27_perm_0 = const()[name = string("value_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199106048)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2428_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = string("op_2428_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199108160)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2430_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = string("op_2430_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_19_perm_0 = const()[name = string("q_with_bias_v_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_241_transpose_x_0 = const()[name = string("x_241_transpose_x_0"), val = bool(false)];
+            bool x_241_transpose_y_0 = const()[name = string("x_241_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_2432_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199110272))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199310016))))[name = string("op_2432_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2430_cast_fp16)[name = string("transpose_281")];
+            tensor<fp16, [1, 8, 56, 195]> x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = op_2432_to_fp16_quantized)[name = string("x_241_cast_fp16")];
+            tensor<int32, [8]> x_243_pad_0 = const()[name = string("x_243_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_243_mode_0 = const()[name = string("x_243_mode_0"), val = string("constant")];
+            fp16 const_196_to_fp16 = const()[name = string("const_196_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_243_cast_fp16 = pad(constant_val = const_196_to_fp16, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241_cast_fp16)[name = string("x_243_cast_fp16")];
+            tensor<int32, [4]> var_2440 = const()[name = string("op_2440"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_245_cast_fp16 = reshape(shape = var_2440, x = x_243_cast_fp16)[name = string("x_245_cast_fp16")];
+            tensor<int32, [4]> var_2444_begin_0 = const()[name = string("op_2444_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2444_end_0 = const()[name = string("op_2444_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_2444_end_mask_0 = const()[name = string("op_2444_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_2444_cast_fp16 = slice_by_index(begin = var_2444_begin_0, end = var_2444_end_0, end_mask = var_2444_end_mask_0, x = x_245_cast_fp16)[name = string("op_2444_cast_fp16")];
+            tensor<int32, [4]> var_2445 = const()[name = string("op_2445"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_37_cast_fp16 = reshape(shape = var_2445, x = var_2444_cast_fp16)[name = string("matrix_bd_37_cast_fp16")];
+            bool matrix_ac_19_transpose_x_0 = const()[name = string("matrix_ac_19_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_19_transpose_y_0 = const()[name = string("matrix_ac_19_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_115 = transpose(perm = transpose_115_perm_0, x = k_37_cast_fp16)[name = string("transpose_279")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_114 = transpose(perm = transpose_114_perm_0, x = var_2428_cast_fp16)[name = string("transpose_280")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_19_cast_fp16 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_114, y = transpose_115)[name = string("matrix_ac_19_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_39_begin_0 = const()[name = string("matrix_bd_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_39_end_0 = const()[name = string("matrix_bd_39_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_39_end_mask_0 = const()[name = string("matrix_bd_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_39_cast_fp16 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37_cast_fp16)[name = string("matrix_bd_39_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2454_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = string("op_2454_cast_fp16")];
+            fp16 _inversed_scores_37_y_0_to_fp16 = const()[name = string("_inversed_scores_37_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_37_cast_fp16 = mul(x = var_2454_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = string("_inversed_scores_37_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_39_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_11)[name = string("scores_39_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2460_cast_fp16 = softmax(axis = var_58, x = scores_39_cast_fp16)[name = string("op_2460_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_509_cast_fp16 = select(a = var_43_to_fp16, b = var_2460_cast_fp16, cond = mask_11)[name = string("input_509_cast_fp16")];
+            bool x_247_transpose_x_0 = const()[name = string("x_247_transpose_x_0"), val = bool(false)];
+            bool x_247_transpose_y_0 = const()[name = string("x_247_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_19_cast_fp16)[name = string("transpose_278")];
+            tensor<fp16, [1, 8, 56, 128]> x_247_cast_fp16 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509_cast_fp16, y = value_27_cast_fp16)[name = string("x_247_cast_fp16")];
+            tensor<int32, [4]> var_2464_perm_0 = const()[name = string("op_2464_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2465 = const()[name = string("op_2465"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_2464_cast_fp16 = transpose(perm = var_2464_perm_0, x = x_247_cast_fp16)[name = string("transpose_277")];
+            tensor<fp16, [1, 56, 1024]> input_511_cast_fp16 = reshape(shape = var_2465, x = var_2464_cast_fp16)[name = string("input_511_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199310528))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200097024))))[name = string("encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200097216)))];
+            tensor<fp16, [1, 56, 1024]> linear_88_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized, x = input_511_cast_fp16)[name = string("linear_88_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = string("input_515_cast_fp16")];
+            tensor<int32, [1]> x_251_axes_0 = const()[name = string("x_251_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200099328)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200101440)))];
+            tensor<fp16, [1, 56, 1024]> x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = string("x_251_cast_fp16")];
+            tensor<int32, [3]> input_517_perm_0 = const()[name = string("input_517_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_519_pad_type_0 = const()[name = string("input_519_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_519_strides_0 = const()[name = string("input_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_519_pad_0 = const()[name = string("input_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_519_dilations_0 = const()[name = string("input_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_519_groups_0 = const()[name = string("input_519_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(200103552))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202200768))))[name = string("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = string("transpose_276")];
+            tensor<fp16, [1, 2048, 56]> input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_517_cast_fp16)[name = string("input_519_cast_fp16")];
+            int32 x_253_split_num_splits_0 = const()[name = string("x_253_split_num_splits_0"), val = int32(2)];
+            int32 x_253_split_axis_0 = const()[name = string("x_253_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_253_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_253_split_cast_fp16_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519_cast_fp16)[name = string("x_253_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = string("x_253_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = string("x_253_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_521_cast_fp16 = select(a = var_43_to_fp16, b = x_253_cast_fp16, cond = var_574)[name = string("input_521_cast_fp16")];
+            bool new_x_39_interleave_0 = const()[name = string("new_x_39_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_39_cast_fp16 = concat(axis = var_58, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = string("new_x_39_cast_fp16")];
+            tensor<int32, [3]> var_2504_begin_0 = const()[name = string("op_2504_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_2504_end_0 = const()[name = string("op_2504_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_2504_end_mask_0 = const()[name = string("op_2504_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2504_cast_fp16 = slice_by_index(begin = var_2504_begin_0, end = var_2504_end_0, end_mask = var_2504_end_mask_0, x = new_x_39_cast_fp16)[name = string("op_2504_cast_fp16")];
+            string x_255_pad_type_0 = const()[name = string("x_255_pad_type_0"), val = string("valid")];
+            int32 x_255_groups_0 = const()[name = string("x_255_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_255_strides_0 = const()[name = string("x_255_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_255_pad_0 = const()[name = string("x_255_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_255_dilations_0 = const()[name = string("x_255_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202204928))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202214208))))[name = string("encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_39_cast_fp16)[name = string("x_255_cast_fp16")];
+            tensor<int32, [3]> input_523_perm_0 = const()[name = string("input_523_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_257_axes_0 = const()[name = string("x_257_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202216320)))];
+            tensor<fp16, [1024]> encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202218432)))];
+            tensor<fp16, [1, 56, 1024]> input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = string("transpose_275")];
+            tensor<fp16, [1, 56, 1024]> x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = string("x_257_cast_fp16")];
+            tensor<int32, [3]> input_525_perm_0 = const()[name = string("input_525_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = string("transpose_274")];
+            tensor<fp16, [1, 1024, 56]> input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = string("input_527_cast_fp16")];
+            string x_259_pad_type_0 = const()[name = string("x_259_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_259_strides_0 = const()[name = string("x_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_259_pad_0 = const()[name = string("x_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_259_dilations_0 = const()[name = string("x_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_259_groups_0 = const()[name = string("x_259_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202220544))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203269184))))[name = string("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_527_cast_fp16)[name = string("x_259_cast_fp16")];
+            tensor<int32, [3]> input_529_perm_0 = const()[name = string("input_529_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = string("transpose_273")];
+            tensor<fp16, [1, 56, 1024]> input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = string("input_531_cast_fp16")];
+            tensor<int32, [1]> input_533_axes_0 = const()[name = string("input_533_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203271296)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203273408)))];
+            tensor<fp16, [1, 56, 1024]> input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = string("input_533_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203275520))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206421312))))[name = string("encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_9_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206421504)))];
+            tensor<fp16, [1, 56, 4096]> linear_89_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized, x = input_533_cast_fp16)[name = string("linear_89_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = string("input_537_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(206429760))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209575552))))[name = string("encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209575744)))];
+            tensor<fp16, [1, 56, 1024]> linear_90_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized, x = input_537_cast_fp16)[name = string("linear_90_cast_fp16")];
+            fp16 var_2547_to_fp16 = const()[name = string("op_2547_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_2548_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2547_to_fp16)[name = string("op_2548_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2548_cast_fp16)[name = string("input_543_cast_fp16")];
+            tensor<int32, [1]> input_545_axes_0 = const()[name = string("input_545_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209577856)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209579968)))];
+            tensor<fp16, [1, 56, 1024]> input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = string("input_545_cast_fp16")];
+            tensor<int32, [4]> cache_41_begin_0 = const()[name = string("cache_41_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_41_end_0 = const()[name = string("cache_41_end_0"), val = tensor<int32, [4]>([11, 1, 42, 1024])];
+            tensor<bool, [4]> cache_41_end_mask_0 = const()[name = string("cache_41_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_41_squeeze_mask_0 = const()[name = string("cache_41_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_41_cast_fp16 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_41_cast_fp16")];
+            tensor<int32, [4]> cache_43_begin_0 = const()[name = string("cache_43_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_43_end_0 = const()[name = string("cache_43_end_0"), val = tensor<int32, [4]>([11, 1, 1024, 8])];
+            tensor<bool, [4]> cache_43_end_mask_0 = const()[name = string("cache_43_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_43_squeeze_mask_0 = const()[name = string("cache_43_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_43_cast_fp16 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_43_cast_fp16")];
+            tensor<int32, [1]> input_547_axes_0 = const()[name = string("input_547_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209582080)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209584192)))];
+            tensor<fp16, [1, 56, 1024]> input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = string("input_547_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209586304))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212732096))))[name = string("encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_10_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212732288)))];
+            tensor<fp16, [1, 56, 4096]> linear_91_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized, x = input_547_cast_fp16)[name = string("linear_91_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = string("input_551_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212740544))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215886336))))[name = string("encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215886528)))];
+            tensor<fp16, [1, 56, 1024]> linear_92_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized, x = input_551_cast_fp16)[name = string("linear_92_cast_fp16")];
+            fp16 var_2584_to_fp16 = const()[name = string("op_2584_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_2585_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2584_to_fp16)[name = string("op_2585_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2585_cast_fp16)[name = string("input_557_cast_fp16")];
+            tensor<int32, [1]> key_21_axes_0 = const()[name = string("key_21_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215888640)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215890752)))];
+            tensor<fp16, [1, 56, 1024]> key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = string("key_21_cast_fp16")];
+            bool input_559_interleave_0 = const()[name = string("input_559_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_559_cast_fp16 = concat(axis = var_67, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = string("input_559_cast_fp16")];
+            bool var_2613_interleave_0 = const()[name = string("op_2613_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_2613_cast_fp16 = concat(axis = var_67, interleave = var_2613_interleave_0, values = key_21_cast_fp16)[name = string("op_2613_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215892864))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216679360))))[name = string("encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216679552)))];
+            tensor<fp16, [1, 56, 1024]> linear_93_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized, x = key_21_cast_fp16)[name = string("linear_93_cast_fp16")];
+            tensor<int32, [4]> var_2618 = const()[name = string("op_2618"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_61_cast_fp16 = reshape(shape = var_2618, x = linear_93_cast_fp16)[name = string("q_61_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216681664))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217468160))))[name = string("encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217468352)))];
+            tensor<fp16, [1, 98, 1024]> linear_94_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_94_cast_fp16")];
+            tensor<int32, [4]> var_2623 = const()[name = string("op_2623"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_41_cast_fp16 = reshape(shape = var_2623, x = linear_94_cast_fp16)[name = string("k_41_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217470464))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218256960))))[name = string("encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218257152)))];
+            tensor<fp16, [1, 98, 1024]> linear_95_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_95_cast_fp16")];
+            tensor<int32, [4]> var_2628 = const()[name = string("op_2628"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_21_cast_fp16 = reshape(shape = var_2628, x = linear_95_cast_fp16)[name = string("v_21_cast_fp16")];
+            tensor<int32, [4]> value_29_perm_0 = const()[name = string("value_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218259264)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2641_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = string("op_2641_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218261376)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2643_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = string("op_2643_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_21_perm_0 = const()[name = string("q_with_bias_v_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_267_transpose_x_0 = const()[name = string("x_267_transpose_x_0"), val = bool(false)];
+            bool x_267_transpose_y_0 = const()[name = string("x_267_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_2645_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218263488))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218463232))))[name = string("op_2645_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2643_cast_fp16)[name = string("transpose_272")];
+            tensor<fp16, [1, 8, 56, 195]> x_267_cast_fp16 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = op_2645_to_fp16_quantized)[name = string("x_267_cast_fp16")];
+            tensor<int32, [8]> x_269_pad_0 = const()[name = string("x_269_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_269_mode_0 = const()[name = string("x_269_mode_0"), val = string("constant")];
+            fp16 const_209_to_fp16 = const()[name = string("const_209_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_269_cast_fp16 = pad(constant_val = const_209_to_fp16, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267_cast_fp16)[name = string("x_269_cast_fp16")];
+            tensor<int32, [4]> var_2653 = const()[name = string("op_2653"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_271_cast_fp16 = reshape(shape = var_2653, x = x_269_cast_fp16)[name = string("x_271_cast_fp16")];
+            tensor<int32, [4]> var_2657_begin_0 = const()[name = string("op_2657_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2657_end_0 = const()[name = string("op_2657_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_2657_end_mask_0 = const()[name = string("op_2657_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_2657_cast_fp16 = slice_by_index(begin = var_2657_begin_0, end = var_2657_end_0, end_mask = var_2657_end_mask_0, x = x_271_cast_fp16)[name = string("op_2657_cast_fp16")];
+            tensor<int32, [4]> var_2658 = const()[name = string("op_2658"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_41_cast_fp16 = reshape(shape = var_2658, x = var_2657_cast_fp16)[name = string("matrix_bd_41_cast_fp16")];
+            bool matrix_ac_21_transpose_x_0 = const()[name = string("matrix_ac_21_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_21_transpose_y_0 = const()[name = string("matrix_ac_21_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_117 = transpose(perm = transpose_117_perm_0, x = k_41_cast_fp16)[name = string("transpose_270")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_116 = transpose(perm = transpose_116_perm_0, x = var_2641_cast_fp16)[name = string("transpose_271")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_21_cast_fp16 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_116, y = transpose_117)[name = string("matrix_ac_21_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_43_begin_0 = const()[name = string("matrix_bd_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_43_end_0 = const()[name = string("matrix_bd_43_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_43_end_mask_0 = const()[name = string("matrix_bd_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_43_cast_fp16 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41_cast_fp16)[name = string("matrix_bd_43_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2667_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = string("op_2667_cast_fp16")];
+            fp16 _inversed_scores_41_y_0_to_fp16 = const()[name = string("_inversed_scores_41_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_41_cast_fp16 = mul(x = var_2667_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = string("_inversed_scores_41_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_43_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_11)[name = string("scores_43_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2673_cast_fp16 = softmax(axis = var_58, x = scores_43_cast_fp16)[name = string("op_2673_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_561_cast_fp16 = select(a = var_43_to_fp16, b = var_2673_cast_fp16, cond = mask_11)[name = string("input_561_cast_fp16")];
+            bool x_273_transpose_x_0 = const()[name = string("x_273_transpose_x_0"), val = bool(false)];
+            bool x_273_transpose_y_0 = const()[name = string("x_273_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_21_cast_fp16)[name = string("transpose_269")];
+            tensor<fp16, [1, 8, 56, 128]> x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561_cast_fp16, y = value_29_cast_fp16)[name = string("x_273_cast_fp16")];
+            tensor<int32, [4]> var_2677_perm_0 = const()[name = string("op_2677_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2678 = const()[name = string("op_2678"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_2677_cast_fp16 = transpose(perm = var_2677_perm_0, x = x_273_cast_fp16)[name = string("transpose_268")];
+            tensor<fp16, [1, 56, 1024]> input_563_cast_fp16 = reshape(shape = var_2678, x = var_2677_cast_fp16)[name = string("input_563_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218463744))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219250240))))[name = string("encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219250432)))];
+            tensor<fp16, [1, 56, 1024]> linear_97_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized, x = input_563_cast_fp16)[name = string("linear_97_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = string("input_567_cast_fp16")];
+            tensor<int32, [1]> x_277_axes_0 = const()[name = string("x_277_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219252544)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219254656)))];
+            tensor<fp16, [1, 56, 1024]> x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = string("x_277_cast_fp16")];
+            tensor<int32, [3]> input_569_perm_0 = const()[name = string("input_569_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_571_pad_type_0 = const()[name = string("input_571_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_571_strides_0 = const()[name = string("input_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_571_pad_0 = const()[name = string("input_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_571_dilations_0 = const()[name = string("input_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_571_groups_0 = const()[name = string("input_571_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(219256768))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221353984))))[name = string("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = string("transpose_267")];
+            tensor<fp16, [1, 2048, 56]> input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_569_cast_fp16)[name = string("input_571_cast_fp16")];
+            int32 x_279_split_num_splits_0 = const()[name = string("x_279_split_num_splits_0"), val = int32(2)];
+            int32 x_279_split_axis_0 = const()[name = string("x_279_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_279_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_279_split_cast_fp16_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571_cast_fp16)[name = string("x_279_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = string("x_279_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = string("x_279_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_573_cast_fp16 = select(a = var_43_to_fp16, b = x_279_cast_fp16, cond = var_574)[name = string("input_573_cast_fp16")];
+            bool new_x_43_interleave_0 = const()[name = string("new_x_43_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_43_cast_fp16 = concat(axis = var_58, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = string("new_x_43_cast_fp16")];
+            tensor<int32, [3]> var_2717_begin_0 = const()[name = string("op_2717_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_2717_end_0 = const()[name = string("op_2717_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_2717_end_mask_0 = const()[name = string("op_2717_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2717_cast_fp16 = slice_by_index(begin = var_2717_begin_0, end = var_2717_end_0, end_mask = var_2717_end_mask_0, x = new_x_43_cast_fp16)[name = string("op_2717_cast_fp16")];
+            string x_281_pad_type_0 = const()[name = string("x_281_pad_type_0"), val = string("valid")];
+            int32 x_281_groups_0 = const()[name = string("x_281_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_281_strides_0 = const()[name = string("x_281_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_281_pad_0 = const()[name = string("x_281_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_281_dilations_0 = const()[name = string("x_281_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221358144))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221367424))))[name = string("encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_281_cast_fp16 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_43_cast_fp16)[name = string("x_281_cast_fp16")];
+            tensor<int32, [3]> input_575_perm_0 = const()[name = string("input_575_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_283_axes_0 = const()[name = string("x_283_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221369536)))];
+            tensor<fp16, [1024]> encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221371648)))];
+            tensor<fp16, [1, 56, 1024]> input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = string("transpose_266")];
+            tensor<fp16, [1, 56, 1024]> x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = string("x_283_cast_fp16")];
+            tensor<int32, [3]> input_577_perm_0 = const()[name = string("input_577_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = string("transpose_265")];
+            tensor<fp16, [1, 1024, 56]> input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = string("input_579_cast_fp16")];
+            string x_285_pad_type_0 = const()[name = string("x_285_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_285_strides_0 = const()[name = string("x_285_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_285_pad_0 = const()[name = string("x_285_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_285_dilations_0 = const()[name = string("x_285_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_285_groups_0 = const()[name = string("x_285_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221373760))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222422400))))[name = string("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_579_cast_fp16)[name = string("x_285_cast_fp16")];
+            tensor<int32, [3]> input_581_perm_0 = const()[name = string("input_581_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = string("transpose_264")];
+            tensor<fp16, [1, 56, 1024]> input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = string("input_583_cast_fp16")];
+            tensor<int32, [1]> input_585_axes_0 = const()[name = string("input_585_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222424512)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222426624)))];
+            tensor<fp16, [1, 56, 1024]> input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = string("input_585_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222428736))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225574528))))[name = string("encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_10_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225574720)))];
+            tensor<fp16, [1, 56, 4096]> linear_98_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized, x = input_585_cast_fp16)[name = string("linear_98_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = string("input_589_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225582976))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228728768))))[name = string("encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228728960)))];
+            tensor<fp16, [1, 56, 1024]> linear_99_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized, x = input_589_cast_fp16)[name = string("linear_99_cast_fp16")];
+            fp16 var_2760_to_fp16 = const()[name = string("op_2760_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_2761_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2760_to_fp16)[name = string("op_2761_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2761_cast_fp16)[name = string("input_595_cast_fp16")];
+            tensor<int32, [1]> input_597_axes_0 = const()[name = string("input_597_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228731072)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228733184)))];
+            tensor<fp16, [1, 56, 1024]> input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = string("input_597_cast_fp16")];
+            tensor<int32, [4]> cache_45_begin_0 = const()[name = string("cache_45_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_45_end_0 = const()[name = string("cache_45_end_0"), val = tensor<int32, [4]>([12, 1, 42, 1024])];
+            tensor<bool, [4]> cache_45_end_mask_0 = const()[name = string("cache_45_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_45_squeeze_mask_0 = const()[name = string("cache_45_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_45_cast_fp16 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_45_cast_fp16")];
+            tensor<int32, [4]> cache_47_begin_0 = const()[name = string("cache_47_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_47_end_0 = const()[name = string("cache_47_end_0"), val = tensor<int32, [4]>([12, 1, 1024, 8])];
+            tensor<bool, [4]> cache_47_end_mask_0 = const()[name = string("cache_47_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_47_squeeze_mask_0 = const()[name = string("cache_47_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_47_cast_fp16 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_47_cast_fp16")];
+            tensor<int32, [1]> input_599_axes_0 = const()[name = string("input_599_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228735296)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228737408)))];
+            tensor<fp16, [1, 56, 1024]> input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = string("input_599_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228739520))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231885312))))[name = string("encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_11_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231885504)))];
+            tensor<fp16, [1, 56, 4096]> linear_100_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized, x = input_599_cast_fp16)[name = string("linear_100_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = string("input_603_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(231893760))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235039552))))[name = string("encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235039744)))];
+            tensor<fp16, [1, 56, 1024]> linear_101_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized, x = input_603_cast_fp16)[name = string("linear_101_cast_fp16")];
+            fp16 var_2797_to_fp16 = const()[name = string("op_2797_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_2798_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2797_to_fp16)[name = string("op_2798_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2798_cast_fp16)[name = string("input_609_cast_fp16")];
+            tensor<int32, [1]> key_23_axes_0 = const()[name = string("key_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235041856)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235043968)))];
+            tensor<fp16, [1, 56, 1024]> key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = string("key_23_cast_fp16")];
+            bool input_611_interleave_0 = const()[name = string("input_611_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_611_cast_fp16 = concat(axis = var_67, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = string("input_611_cast_fp16")];
+            bool var_2826_interleave_0 = const()[name = string("op_2826_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_2826_cast_fp16 = concat(axis = var_67, interleave = var_2826_interleave_0, values = key_23_cast_fp16)[name = string("op_2826_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235046080))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235832576))))[name = string("encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235832768)))];
+            tensor<fp16, [1, 56, 1024]> linear_102_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized, x = key_23_cast_fp16)[name = string("linear_102_cast_fp16")];
+            tensor<int32, [4]> var_2831 = const()[name = string("op_2831"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_67_cast_fp16 = reshape(shape = var_2831, x = linear_102_cast_fp16)[name = string("q_67_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235834880))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236621376))))[name = string("encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236621568)))];
+            tensor<fp16, [1, 98, 1024]> linear_103_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_103_cast_fp16")];
+            tensor<int32, [4]> var_2836 = const()[name = string("op_2836"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_45_cast_fp16 = reshape(shape = var_2836, x = linear_103_cast_fp16)[name = string("k_45_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236623680))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237410176))))[name = string("encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237410368)))];
+            tensor<fp16, [1, 98, 1024]> linear_104_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_104_cast_fp16")];
+            tensor<int32, [4]> var_2841 = const()[name = string("op_2841"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_23_cast_fp16 = reshape(shape = var_2841, x = linear_104_cast_fp16)[name = string("v_23_cast_fp16")];
+            tensor<int32, [4]> value_31_perm_0 = const()[name = string("value_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237412480)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2854_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = string("op_2854_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237414592)))];
+            tensor<fp16, [1, 56, 8, 128]> var_2856_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = string("op_2856_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_23_perm_0 = const()[name = string("q_with_bias_v_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_293_transpose_x_0 = const()[name = string("x_293_transpose_x_0"), val = bool(false)];
+            bool x_293_transpose_y_0 = const()[name = string("x_293_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_2858_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237416704))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237616448))))[name = string("op_2858_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2856_cast_fp16)[name = string("transpose_263")];
+            tensor<fp16, [1, 8, 56, 195]> x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = op_2858_to_fp16_quantized)[name = string("x_293_cast_fp16")];
+            tensor<int32, [8]> x_295_pad_0 = const()[name = string("x_295_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_295_mode_0 = const()[name = string("x_295_mode_0"), val = string("constant")];
+            fp16 const_222_to_fp16 = const()[name = string("const_222_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_295_cast_fp16 = pad(constant_val = const_222_to_fp16, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293_cast_fp16)[name = string("x_295_cast_fp16")];
+            tensor<int32, [4]> var_2866 = const()[name = string("op_2866"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_297_cast_fp16 = reshape(shape = var_2866, x = x_295_cast_fp16)[name = string("x_297_cast_fp16")];
+            tensor<int32, [4]> var_2870_begin_0 = const()[name = string("op_2870_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2870_end_0 = const()[name = string("op_2870_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_2870_end_mask_0 = const()[name = string("op_2870_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_2870_cast_fp16 = slice_by_index(begin = var_2870_begin_0, end = var_2870_end_0, end_mask = var_2870_end_mask_0, x = x_297_cast_fp16)[name = string("op_2870_cast_fp16")];
+            tensor<int32, [4]> var_2871 = const()[name = string("op_2871"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_45_cast_fp16 = reshape(shape = var_2871, x = var_2870_cast_fp16)[name = string("matrix_bd_45_cast_fp16")];
+            bool matrix_ac_23_transpose_x_0 = const()[name = string("matrix_ac_23_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_23_transpose_y_0 = const()[name = string("matrix_ac_23_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_119 = transpose(perm = transpose_119_perm_0, x = k_45_cast_fp16)[name = string("transpose_261")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_118 = transpose(perm = transpose_118_perm_0, x = var_2854_cast_fp16)[name = string("transpose_262")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_23_cast_fp16 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_118, y = transpose_119)[name = string("matrix_ac_23_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_47_begin_0 = const()[name = string("matrix_bd_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_47_end_0 = const()[name = string("matrix_bd_47_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_47_end_mask_0 = const()[name = string("matrix_bd_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_47_cast_fp16 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45_cast_fp16)[name = string("matrix_bd_47_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2880_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = string("op_2880_cast_fp16")];
+            fp16 _inversed_scores_45_y_0_to_fp16 = const()[name = string("_inversed_scores_45_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_45_cast_fp16 = mul(x = var_2880_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = string("_inversed_scores_45_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_47_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_11)[name = string("scores_47_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_2886_cast_fp16 = softmax(axis = var_58, x = scores_47_cast_fp16)[name = string("op_2886_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_613_cast_fp16 = select(a = var_43_to_fp16, b = var_2886_cast_fp16, cond = mask_11)[name = string("input_613_cast_fp16")];
+            bool x_299_transpose_x_0 = const()[name = string("x_299_transpose_x_0"), val = bool(false)];
+            bool x_299_transpose_y_0 = const()[name = string("x_299_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_23_cast_fp16)[name = string("transpose_260")];
+            tensor<fp16, [1, 8, 56, 128]> x_299_cast_fp16 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613_cast_fp16, y = value_31_cast_fp16)[name = string("x_299_cast_fp16")];
+            tensor<int32, [4]> var_2890_perm_0 = const()[name = string("op_2890_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2891 = const()[name = string("op_2891"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_2890_cast_fp16 = transpose(perm = var_2890_perm_0, x = x_299_cast_fp16)[name = string("transpose_259")];
+            tensor<fp16, [1, 56, 1024]> input_615_cast_fp16 = reshape(shape = var_2891, x = var_2890_cast_fp16)[name = string("input_615_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237616960))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238403456))))[name = string("encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238403648)))];
+            tensor<fp16, [1, 56, 1024]> linear_106_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized, x = input_615_cast_fp16)[name = string("linear_106_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = string("input_619_cast_fp16")];
+            tensor<int32, [1]> x_303_axes_0 = const()[name = string("x_303_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238405760)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238407872)))];
+            tensor<fp16, [1, 56, 1024]> x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = string("x_303_cast_fp16")];
+            tensor<int32, [3]> input_621_perm_0 = const()[name = string("input_621_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_623_pad_type_0 = const()[name = string("input_623_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_623_strides_0 = const()[name = string("input_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_623_pad_0 = const()[name = string("input_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_623_dilations_0 = const()[name = string("input_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_623_groups_0 = const()[name = string("input_623_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(238409984))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240507200))))[name = string("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = string("transpose_258")];
+            tensor<fp16, [1, 2048, 56]> input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_621_cast_fp16)[name = string("input_623_cast_fp16")];
+            int32 x_305_split_num_splits_0 = const()[name = string("x_305_split_num_splits_0"), val = int32(2)];
+            int32 x_305_split_axis_0 = const()[name = string("x_305_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_305_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623_cast_fp16)[name = string("x_305_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = string("x_305_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = string("x_305_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_625_cast_fp16 = select(a = var_43_to_fp16, b = x_305_cast_fp16, cond = var_574)[name = string("input_625_cast_fp16")];
+            bool new_x_47_interleave_0 = const()[name = string("new_x_47_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_47_cast_fp16 = concat(axis = var_58, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = string("new_x_47_cast_fp16")];
+            tensor<int32, [3]> var_2930_begin_0 = const()[name = string("op_2930_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_2930_end_0 = const()[name = string("op_2930_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_2930_end_mask_0 = const()[name = string("op_2930_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2930_cast_fp16 = slice_by_index(begin = var_2930_begin_0, end = var_2930_end_0, end_mask = var_2930_end_mask_0, x = new_x_47_cast_fp16)[name = string("op_2930_cast_fp16")];
+            string x_307_pad_type_0 = const()[name = string("x_307_pad_type_0"), val = string("valid")];
+            int32 x_307_groups_0 = const()[name = string("x_307_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_307_strides_0 = const()[name = string("x_307_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_307_pad_0 = const()[name = string("x_307_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_307_dilations_0 = const()[name = string("x_307_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240511360))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240520640))))[name = string("encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_47_cast_fp16)[name = string("x_307_cast_fp16")];
+            tensor<int32, [3]> input_627_perm_0 = const()[name = string("input_627_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_309_axes_0 = const()[name = string("x_309_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240522752)))];
+            tensor<fp16, [1024]> encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240524864)))];
+            tensor<fp16, [1, 56, 1024]> input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = string("transpose_257")];
+            tensor<fp16, [1, 56, 1024]> x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = string("x_309_cast_fp16")];
+            tensor<int32, [3]> input_629_perm_0 = const()[name = string("input_629_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = string("transpose_256")];
+            tensor<fp16, [1, 1024, 56]> input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = string("input_631_cast_fp16")];
+            string x_311_pad_type_0 = const()[name = string("x_311_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_311_strides_0 = const()[name = string("x_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_311_pad_0 = const()[name = string("x_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_311_dilations_0 = const()[name = string("x_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_311_groups_0 = const()[name = string("x_311_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240526976))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241575616))))[name = string("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_311_cast_fp16 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_631_cast_fp16)[name = string("x_311_cast_fp16")];
+            tensor<int32, [3]> input_633_perm_0 = const()[name = string("input_633_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = string("transpose_255")];
+            tensor<fp16, [1, 56, 1024]> input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = string("input_635_cast_fp16")];
+            tensor<int32, [1]> input_637_axes_0 = const()[name = string("input_637_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241577728)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241579840)))];
+            tensor<fp16, [1, 56, 1024]> input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = string("input_637_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241581952))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244727744))))[name = string("encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_11_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244727936)))];
+            tensor<fp16, [1, 56, 4096]> linear_107_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized, x = input_637_cast_fp16)[name = string("linear_107_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = string("input_641_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(244736192))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247881984))))[name = string("encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247882176)))];
+            tensor<fp16, [1, 56, 1024]> linear_108_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized, x = input_641_cast_fp16)[name = string("linear_108_cast_fp16")];
+            fp16 var_2973_to_fp16 = const()[name = string("op_2973_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_2974_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2973_to_fp16)[name = string("op_2974_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2974_cast_fp16)[name = string("input_647_cast_fp16")];
+            tensor<int32, [1]> input_649_axes_0 = const()[name = string("input_649_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247884288)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247886400)))];
+            tensor<fp16, [1, 56, 1024]> input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = string("input_649_cast_fp16")];
+            tensor<int32, [4]> cache_49_begin_0 = const()[name = string("cache_49_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_49_end_0 = const()[name = string("cache_49_end_0"), val = tensor<int32, [4]>([13, 1, 42, 1024])];
+            tensor<bool, [4]> cache_49_end_mask_0 = const()[name = string("cache_49_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_49_squeeze_mask_0 = const()[name = string("cache_49_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_49_cast_fp16 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_49_cast_fp16")];
+            tensor<int32, [4]> cache_51_begin_0 = const()[name = string("cache_51_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_51_end_0 = const()[name = string("cache_51_end_0"), val = tensor<int32, [4]>([13, 1, 1024, 8])];
+            tensor<bool, [4]> cache_51_end_mask_0 = const()[name = string("cache_51_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_51_squeeze_mask_0 = const()[name = string("cache_51_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_51_cast_fp16 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_51_cast_fp16")];
+            tensor<int32, [1]> input_651_axes_0 = const()[name = string("input_651_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247888512)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247890624)))];
+            tensor<fp16, [1, 56, 1024]> input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = string("input_651_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247892736))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251038528))))[name = string("encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_12_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251038720)))];
+            tensor<fp16, [1, 56, 4096]> linear_109_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized, x = input_651_cast_fp16)[name = string("linear_109_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("input_655_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251046976))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254192768))))[name = string("encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254192960)))];
+            tensor<fp16, [1, 56, 1024]> linear_110_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized, x = input_655_cast_fp16)[name = string("linear_110_cast_fp16")];
+            fp16 var_3010_to_fp16 = const()[name = string("op_3010_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_3011_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_3010_to_fp16)[name = string("op_3011_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_3011_cast_fp16)[name = string("input_661_cast_fp16")];
+            tensor<int32, [1]> key_25_axes_0 = const()[name = string("key_25_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254195072)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254197184)))];
+            tensor<fp16, [1, 56, 1024]> key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = string("key_25_cast_fp16")];
+            bool input_663_interleave_0 = const()[name = string("input_663_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_663_cast_fp16 = concat(axis = var_67, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = string("input_663_cast_fp16")];
+            bool var_3039_interleave_0 = const()[name = string("op_3039_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_3039_cast_fp16 = concat(axis = var_67, interleave = var_3039_interleave_0, values = key_25_cast_fp16)[name = string("op_3039_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254199296))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254985792))))[name = string("encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254985984)))];
+            tensor<fp16, [1, 56, 1024]> linear_111_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized, x = key_25_cast_fp16)[name = string("linear_111_cast_fp16")];
+            tensor<int32, [4]> var_3044 = const()[name = string("op_3044"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_73_cast_fp16 = reshape(shape = var_3044, x = linear_111_cast_fp16)[name = string("q_73_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254988096))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255774592))))[name = string("encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255774784)))];
+            tensor<fp16, [1, 98, 1024]> linear_112_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_112_cast_fp16")];
+            tensor<int32, [4]> var_3049 = const()[name = string("op_3049"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_49_cast_fp16 = reshape(shape = var_3049, x = linear_112_cast_fp16)[name = string("k_49_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255776896))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256563392))))[name = string("encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256563584)))];
+            tensor<fp16, [1, 98, 1024]> linear_113_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_113_cast_fp16")];
+            tensor<int32, [4]> var_3054 = const()[name = string("op_3054"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_25_cast_fp16 = reshape(shape = var_3054, x = linear_113_cast_fp16)[name = string("v_25_cast_fp16")];
+            tensor<int32, [4]> value_33_perm_0 = const()[name = string("value_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256565696)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3067_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = string("op_3067_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256567808)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3069_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = string("op_3069_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_25_perm_0 = const()[name = string("q_with_bias_v_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_319_transpose_x_0 = const()[name = string("x_319_transpose_x_0"), val = bool(false)];
+            bool x_319_transpose_y_0 = const()[name = string("x_319_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_3071_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256569920))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256769664))))[name = string("op_3071_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_3069_cast_fp16)[name = string("transpose_254")];
+            tensor<fp16, [1, 8, 56, 195]> x_319_cast_fp16 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = op_3071_to_fp16_quantized)[name = string("x_319_cast_fp16")];
+            tensor<int32, [8]> x_321_pad_0 = const()[name = string("x_321_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("constant")];
+            fp16 const_235_to_fp16 = const()[name = string("const_235_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_321_cast_fp16 = pad(constant_val = const_235_to_fp16, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319_cast_fp16)[name = string("x_321_cast_fp16")];
+            tensor<int32, [4]> var_3079 = const()[name = string("op_3079"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_323_cast_fp16 = reshape(shape = var_3079, x = x_321_cast_fp16)[name = string("x_323_cast_fp16")];
+            tensor<int32, [4]> var_3083_begin_0 = const()[name = string("op_3083_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3083_end_0 = const()[name = string("op_3083_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_3083_end_mask_0 = const()[name = string("op_3083_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_3083_cast_fp16 = slice_by_index(begin = var_3083_begin_0, end = var_3083_end_0, end_mask = var_3083_end_mask_0, x = x_323_cast_fp16)[name = string("op_3083_cast_fp16")];
+            tensor<int32, [4]> var_3084 = const()[name = string("op_3084"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_49_cast_fp16 = reshape(shape = var_3084, x = var_3083_cast_fp16)[name = string("matrix_bd_49_cast_fp16")];
+            bool matrix_ac_25_transpose_x_0 = const()[name = string("matrix_ac_25_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_25_transpose_y_0 = const()[name = string("matrix_ac_25_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_120_perm_0 = const()[name = string("transpose_120_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_121_perm_0 = const()[name = string("transpose_121_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_121 = transpose(perm = transpose_121_perm_0, x = k_49_cast_fp16)[name = string("transpose_252")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_120 = transpose(perm = transpose_120_perm_0, x = var_3067_cast_fp16)[name = string("transpose_253")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_25_cast_fp16 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_120, y = transpose_121)[name = string("matrix_ac_25_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_51_begin_0 = const()[name = string("matrix_bd_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_51_end_0 = const()[name = string("matrix_bd_51_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_51_end_mask_0 = const()[name = string("matrix_bd_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_51_cast_fp16 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49_cast_fp16)[name = string("matrix_bd_51_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3093_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = string("op_3093_cast_fp16")];
+            fp16 _inversed_scores_49_y_0_to_fp16 = const()[name = string("_inversed_scores_49_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_49_cast_fp16 = mul(x = var_3093_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = string("_inversed_scores_49_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_51_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_11)[name = string("scores_51_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3099_cast_fp16 = softmax(axis = var_58, x = scores_51_cast_fp16)[name = string("op_3099_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_665_cast_fp16 = select(a = var_43_to_fp16, b = var_3099_cast_fp16, cond = mask_11)[name = string("input_665_cast_fp16")];
+            bool x_325_transpose_x_0 = const()[name = string("x_325_transpose_x_0"), val = bool(false)];
+            bool x_325_transpose_y_0 = const()[name = string("x_325_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_25_cast_fp16)[name = string("transpose_251")];
+            tensor<fp16, [1, 8, 56, 128]> x_325_cast_fp16 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665_cast_fp16, y = value_33_cast_fp16)[name = string("x_325_cast_fp16")];
+            tensor<int32, [4]> var_3103_perm_0 = const()[name = string("op_3103_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3104 = const()[name = string("op_3104"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_3103_cast_fp16 = transpose(perm = var_3103_perm_0, x = x_325_cast_fp16)[name = string("transpose_250")];
+            tensor<fp16, [1, 56, 1024]> input_667_cast_fp16 = reshape(shape = var_3104, x = var_3103_cast_fp16)[name = string("input_667_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256770176))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257556672))))[name = string("encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257556864)))];
+            tensor<fp16, [1, 56, 1024]> linear_115_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized, x = input_667_cast_fp16)[name = string("linear_115_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = string("input_671_cast_fp16")];
+            tensor<int32, [1]> x_329_axes_0 = const()[name = string("x_329_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257558976)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257561088)))];
+            tensor<fp16, [1, 56, 1024]> x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = string("x_329_cast_fp16")];
+            tensor<int32, [3]> input_673_perm_0 = const()[name = string("input_673_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_675_pad_type_0 = const()[name = string("input_675_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_675_strides_0 = const()[name = string("input_675_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_675_pad_0 = const()[name = string("input_675_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_675_dilations_0 = const()[name = string("input_675_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_675_groups_0 = const()[name = string("input_675_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(257563200))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259660416))))[name = string("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = string("transpose_249")];
+            tensor<fp16, [1, 2048, 56]> input_675_cast_fp16 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_673_cast_fp16)[name = string("input_675_cast_fp16")];
+            int32 x_331_split_num_splits_0 = const()[name = string("x_331_split_num_splits_0"), val = int32(2)];
+            int32 x_331_split_axis_0 = const()[name = string("x_331_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_331_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_331_split_cast_fp16_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675_cast_fp16)[name = string("x_331_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = string("x_331_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = string("x_331_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_677_cast_fp16 = select(a = var_43_to_fp16, b = x_331_cast_fp16, cond = var_574)[name = string("input_677_cast_fp16")];
+            bool new_x_51_interleave_0 = const()[name = string("new_x_51_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_51_cast_fp16 = concat(axis = var_58, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = string("new_x_51_cast_fp16")];
+            tensor<int32, [3]> var_3143_begin_0 = const()[name = string("op_3143_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_3143_end_0 = const()[name = string("op_3143_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_3143_end_mask_0 = const()[name = string("op_3143_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3143_cast_fp16 = slice_by_index(begin = var_3143_begin_0, end = var_3143_end_0, end_mask = var_3143_end_mask_0, x = new_x_51_cast_fp16)[name = string("op_3143_cast_fp16")];
+            string x_333_pad_type_0 = const()[name = string("x_333_pad_type_0"), val = string("valid")];
+            int32 x_333_groups_0 = const()[name = string("x_333_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_333_strides_0 = const()[name = string("x_333_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_333_pad_0 = const()[name = string("x_333_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_333_dilations_0 = const()[name = string("x_333_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259664576))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259673856))))[name = string("encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_333_cast_fp16 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_51_cast_fp16)[name = string("x_333_cast_fp16")];
+            tensor<int32, [3]> input_679_perm_0 = const()[name = string("input_679_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_335_axes_0 = const()[name = string("x_335_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259675968)))];
+            tensor<fp16, [1024]> encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259678080)))];
+            tensor<fp16, [1, 56, 1024]> input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = string("transpose_248")];
+            tensor<fp16, [1, 56, 1024]> x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = string("x_335_cast_fp16")];
+            tensor<int32, [3]> input_681_perm_0 = const()[name = string("input_681_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = string("transpose_247")];
+            tensor<fp16, [1, 1024, 56]> input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = string("input_683_cast_fp16")];
+            string x_337_pad_type_0 = const()[name = string("x_337_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_337_strides_0 = const()[name = string("x_337_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_337_pad_0 = const()[name = string("x_337_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_337_dilations_0 = const()[name = string("x_337_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_337_groups_0 = const()[name = string("x_337_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259680192))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260728832))))[name = string("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_337_cast_fp16 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_683_cast_fp16)[name = string("x_337_cast_fp16")];
+            tensor<int32, [3]> input_685_perm_0 = const()[name = string("input_685_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = string("transpose_246")];
+            tensor<fp16, [1, 56, 1024]> input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = string("input_687_cast_fp16")];
+            tensor<int32, [1]> input_689_axes_0 = const()[name = string("input_689_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260730944)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260733056)))];
+            tensor<fp16, [1, 56, 1024]> input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = string("input_689_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260735168))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263880960))))[name = string("encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_12_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263881152)))];
+            tensor<fp16, [1, 56, 4096]> linear_116_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized, x = input_689_cast_fp16)[name = string("linear_116_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("input_693_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263889408))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267035200))))[name = string("encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267035392)))];
+            tensor<fp16, [1, 56, 1024]> linear_117_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized, x = input_693_cast_fp16)[name = string("linear_117_cast_fp16")];
+            fp16 var_3186_to_fp16 = const()[name = string("op_3186_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_3187_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_3186_to_fp16)[name = string("op_3187_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_3187_cast_fp16)[name = string("input_699_cast_fp16")];
+            tensor<int32, [1]> input_701_axes_0 = const()[name = string("input_701_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267037504)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267039616)))];
+            tensor<fp16, [1, 56, 1024]> input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = string("input_701_cast_fp16")];
+            tensor<int32, [4]> cache_53_begin_0 = const()[name = string("cache_53_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_53_end_0 = const()[name = string("cache_53_end_0"), val = tensor<int32, [4]>([14, 1, 42, 1024])];
+            tensor<bool, [4]> cache_53_end_mask_0 = const()[name = string("cache_53_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_53_squeeze_mask_0 = const()[name = string("cache_53_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_53_cast_fp16 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_53_cast_fp16")];
+            tensor<int32, [4]> cache_55_begin_0 = const()[name = string("cache_55_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_55_end_0 = const()[name = string("cache_55_end_0"), val = tensor<int32, [4]>([14, 1, 1024, 8])];
+            tensor<bool, [4]> cache_55_end_mask_0 = const()[name = string("cache_55_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_55_squeeze_mask_0 = const()[name = string("cache_55_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_55_cast_fp16 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_55_cast_fp16")];
+            tensor<int32, [1]> input_703_axes_0 = const()[name = string("input_703_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267041728)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267043840)))];
+            tensor<fp16, [1, 56, 1024]> input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = string("input_703_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(267045952))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270191744))))[name = string("encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_13_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270191936)))];
+            tensor<fp16, [1, 56, 4096]> linear_118_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized, x = input_703_cast_fp16)[name = string("linear_118_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = string("input_707_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270200192))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273345984))))[name = string("encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273346176)))];
+            tensor<fp16, [1, 56, 1024]> linear_119_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized, x = input_707_cast_fp16)[name = string("linear_119_cast_fp16")];
+            fp16 var_3223_to_fp16 = const()[name = string("op_3223_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_3224_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_3223_to_fp16)[name = string("op_3224_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_3224_cast_fp16)[name = string("input_713_cast_fp16")];
+            tensor<int32, [1]> key_27_axes_0 = const()[name = string("key_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273348288)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273350400)))];
+            tensor<fp16, [1, 56, 1024]> key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = string("key_27_cast_fp16")];
+            bool input_715_interleave_0 = const()[name = string("input_715_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_715_cast_fp16 = concat(axis = var_67, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = string("input_715_cast_fp16")];
+            bool var_3252_interleave_0 = const()[name = string("op_3252_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_3252_cast_fp16 = concat(axis = var_67, interleave = var_3252_interleave_0, values = key_27_cast_fp16)[name = string("op_3252_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273352512))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274139008))))[name = string("encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274139200)))];
+            tensor<fp16, [1, 56, 1024]> linear_120_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized, x = key_27_cast_fp16)[name = string("linear_120_cast_fp16")];
+            tensor<int32, [4]> var_3257 = const()[name = string("op_3257"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_79_cast_fp16 = reshape(shape = var_3257, x = linear_120_cast_fp16)[name = string("q_79_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274141312))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274927808))))[name = string("encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274928000)))];
+            tensor<fp16, [1, 98, 1024]> linear_121_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_121_cast_fp16")];
+            tensor<int32, [4]> var_3262 = const()[name = string("op_3262"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_53_cast_fp16 = reshape(shape = var_3262, x = linear_121_cast_fp16)[name = string("k_53_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274930112))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275716608))))[name = string("encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275716800)))];
+            tensor<fp16, [1, 98, 1024]> linear_122_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_122_cast_fp16")];
+            tensor<int32, [4]> var_3267 = const()[name = string("op_3267"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_27_cast_fp16 = reshape(shape = var_3267, x = linear_122_cast_fp16)[name = string("v_27_cast_fp16")];
+            tensor<int32, [4]> value_35_perm_0 = const()[name = string("value_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275718912)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3280_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = string("op_3280_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275721024)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3282_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = string("op_3282_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_27_perm_0 = const()[name = string("q_with_bias_v_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_345_transpose_x_0 = const()[name = string("x_345_transpose_x_0"), val = bool(false)];
+            bool x_345_transpose_y_0 = const()[name = string("x_345_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_3284_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275723136))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275922880))))[name = string("op_3284_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3282_cast_fp16)[name = string("transpose_245")];
+            tensor<fp16, [1, 8, 56, 195]> x_345_cast_fp16 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = op_3284_to_fp16_quantized)[name = string("x_345_cast_fp16")];
+            tensor<int32, [8]> x_347_pad_0 = const()[name = string("x_347_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_347_mode_0 = const()[name = string("x_347_mode_0"), val = string("constant")];
+            fp16 const_248_to_fp16 = const()[name = string("const_248_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_347_cast_fp16 = pad(constant_val = const_248_to_fp16, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345_cast_fp16)[name = string("x_347_cast_fp16")];
+            tensor<int32, [4]> var_3292 = const()[name = string("op_3292"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_349_cast_fp16 = reshape(shape = var_3292, x = x_347_cast_fp16)[name = string("x_349_cast_fp16")];
+            tensor<int32, [4]> var_3296_begin_0 = const()[name = string("op_3296_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3296_end_0 = const()[name = string("op_3296_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_3296_end_mask_0 = const()[name = string("op_3296_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_3296_cast_fp16 = slice_by_index(begin = var_3296_begin_0, end = var_3296_end_0, end_mask = var_3296_end_mask_0, x = x_349_cast_fp16)[name = string("op_3296_cast_fp16")];
+            tensor<int32, [4]> var_3297 = const()[name = string("op_3297"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_53_cast_fp16 = reshape(shape = var_3297, x = var_3296_cast_fp16)[name = string("matrix_bd_53_cast_fp16")];
+            bool matrix_ac_27_transpose_x_0 = const()[name = string("matrix_ac_27_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_27_transpose_y_0 = const()[name = string("matrix_ac_27_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_122_perm_0 = const()[name = string("transpose_122_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_123_perm_0 = const()[name = string("transpose_123_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_123 = transpose(perm = transpose_123_perm_0, x = k_53_cast_fp16)[name = string("transpose_243")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_122 = transpose(perm = transpose_122_perm_0, x = var_3280_cast_fp16)[name = string("transpose_244")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_27_cast_fp16 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_122, y = transpose_123)[name = string("matrix_ac_27_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_55_begin_0 = const()[name = string("matrix_bd_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_55_end_0 = const()[name = string("matrix_bd_55_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_55_end_mask_0 = const()[name = string("matrix_bd_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_55_cast_fp16 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53_cast_fp16)[name = string("matrix_bd_55_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3306_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = string("op_3306_cast_fp16")];
+            fp16 _inversed_scores_53_y_0_to_fp16 = const()[name = string("_inversed_scores_53_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_53_cast_fp16 = mul(x = var_3306_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = string("_inversed_scores_53_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_55_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_11)[name = string("scores_55_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3312_cast_fp16 = softmax(axis = var_58, x = scores_55_cast_fp16)[name = string("op_3312_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_717_cast_fp16 = select(a = var_43_to_fp16, b = var_3312_cast_fp16, cond = mask_11)[name = string("input_717_cast_fp16")];
+            bool x_351_transpose_x_0 = const()[name = string("x_351_transpose_x_0"), val = bool(false)];
+            bool x_351_transpose_y_0 = const()[name = string("x_351_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = v_27_cast_fp16)[name = string("transpose_242")];
+            tensor<fp16, [1, 8, 56, 128]> x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717_cast_fp16, y = value_35_cast_fp16)[name = string("x_351_cast_fp16")];
+            tensor<int32, [4]> var_3316_perm_0 = const()[name = string("op_3316_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3317 = const()[name = string("op_3317"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_3316_cast_fp16 = transpose(perm = var_3316_perm_0, x = x_351_cast_fp16)[name = string("transpose_241")];
+            tensor<fp16, [1, 56, 1024]> input_719_cast_fp16 = reshape(shape = var_3317, x = var_3316_cast_fp16)[name = string("input_719_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275923392))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276709888))))[name = string("encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276710080)))];
+            tensor<fp16, [1, 56, 1024]> linear_124_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("linear_124_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = string("input_723_cast_fp16")];
+            tensor<int32, [1]> x_355_axes_0 = const()[name = string("x_355_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276712192)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276714304)))];
+            tensor<fp16, [1, 56, 1024]> x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = string("x_355_cast_fp16")];
+            tensor<int32, [3]> input_725_perm_0 = const()[name = string("input_725_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_727_pad_type_0 = const()[name = string("input_727_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_727_strides_0 = const()[name = string("input_727_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_727_pad_0 = const()[name = string("input_727_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_727_dilations_0 = const()[name = string("input_727_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_727_groups_0 = const()[name = string("input_727_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276716416))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278813632))))[name = string("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = string("transpose_240")];
+            tensor<fp16, [1, 2048, 56]> input_727_cast_fp16 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_725_cast_fp16)[name = string("input_727_cast_fp16")];
+            int32 x_357_split_num_splits_0 = const()[name = string("x_357_split_num_splits_0"), val = int32(2)];
+            int32 x_357_split_axis_0 = const()[name = string("x_357_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_357_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727_cast_fp16)[name = string("x_357_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = string("x_357_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = string("x_357_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_729_cast_fp16 = select(a = var_43_to_fp16, b = x_357_cast_fp16, cond = var_574)[name = string("input_729_cast_fp16")];
+            bool new_x_55_interleave_0 = const()[name = string("new_x_55_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_55_cast_fp16 = concat(axis = var_58, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = string("new_x_55_cast_fp16")];
+            tensor<int32, [3]> var_3356_begin_0 = const()[name = string("op_3356_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_3356_end_0 = const()[name = string("op_3356_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_3356_end_mask_0 = const()[name = string("op_3356_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3356_cast_fp16 = slice_by_index(begin = var_3356_begin_0, end = var_3356_end_0, end_mask = var_3356_end_mask_0, x = new_x_55_cast_fp16)[name = string("op_3356_cast_fp16")];
+            string x_359_pad_type_0 = const()[name = string("x_359_pad_type_0"), val = string("valid")];
+            int32 x_359_groups_0 = const()[name = string("x_359_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_359_strides_0 = const()[name = string("x_359_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_359_pad_0 = const()[name = string("x_359_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_359_dilations_0 = const()[name = string("x_359_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278817792))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278827072))))[name = string("encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_55_cast_fp16)[name = string("x_359_cast_fp16")];
+            tensor<int32, [3]> input_731_perm_0 = const()[name = string("input_731_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_361_axes_0 = const()[name = string("x_361_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278829184)))];
+            tensor<fp16, [1024]> encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278831296)))];
+            tensor<fp16, [1, 56, 1024]> input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = string("transpose_239")];
+            tensor<fp16, [1, 56, 1024]> x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = string("x_361_cast_fp16")];
+            tensor<int32, [3]> input_733_perm_0 = const()[name = string("input_733_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = string("transpose_238")];
+            tensor<fp16, [1, 1024, 56]> input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = string("input_735_cast_fp16")];
+            string x_363_pad_type_0 = const()[name = string("x_363_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_363_strides_0 = const()[name = string("x_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_363_pad_0 = const()[name = string("x_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_363_dilations_0 = const()[name = string("x_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_363_groups_0 = const()[name = string("x_363_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278833408))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279882048))))[name = string("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_735_cast_fp16)[name = string("x_363_cast_fp16")];
+            tensor<int32, [3]> input_737_perm_0 = const()[name = string("input_737_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = string("transpose_237")];
+            tensor<fp16, [1, 56, 1024]> input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = string("input_739_cast_fp16")];
+            tensor<int32, [1]> input_741_axes_0 = const()[name = string("input_741_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279884160)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279886272)))];
+            tensor<fp16, [1, 56, 1024]> input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = string("input_741_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279888384))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283034176))))[name = string("encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_13_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283034368)))];
+            tensor<fp16, [1, 56, 4096]> linear_125_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized, x = input_741_cast_fp16)[name = string("linear_125_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = string("input_745_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283042624))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286188416))))[name = string("encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286188608)))];
+            tensor<fp16, [1, 56, 1024]> linear_126_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized, x = input_745_cast_fp16)[name = string("linear_126_cast_fp16")];
+            fp16 var_3399_to_fp16 = const()[name = string("op_3399_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_3400_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3399_to_fp16)[name = string("op_3400_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3400_cast_fp16)[name = string("input_751_cast_fp16")];
+            tensor<int32, [1]> input_753_axes_0 = const()[name = string("input_753_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286190720)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286192832)))];
+            tensor<fp16, [1, 56, 1024]> input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = string("input_753_cast_fp16")];
+            tensor<int32, [4]> cache_57_begin_0 = const()[name = string("cache_57_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_57_end_0 = const()[name = string("cache_57_end_0"), val = tensor<int32, [4]>([15, 1, 42, 1024])];
+            tensor<bool, [4]> cache_57_end_mask_0 = const()[name = string("cache_57_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_57_squeeze_mask_0 = const()[name = string("cache_57_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_57_cast_fp16 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_57_cast_fp16")];
+            tensor<int32, [4]> cache_59_begin_0 = const()[name = string("cache_59_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_59_end_0 = const()[name = string("cache_59_end_0"), val = tensor<int32, [4]>([15, 1, 1024, 8])];
+            tensor<bool, [4]> cache_59_end_mask_0 = const()[name = string("cache_59_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_59_squeeze_mask_0 = const()[name = string("cache_59_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_59_cast_fp16 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_59_cast_fp16")];
+            tensor<int32, [1]> input_755_axes_0 = const()[name = string("input_755_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286194944)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286197056)))];
+            tensor<fp16, [1, 56, 1024]> input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = string("input_755_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286199168))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289344960))))[name = string("encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_14_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289345152)))];
+            tensor<fp16, [1, 56, 4096]> linear_127_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized, x = input_755_cast_fp16)[name = string("linear_127_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = string("input_759_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(289353408))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292499200))))[name = string("encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292499392)))];
+            tensor<fp16, [1, 56, 1024]> linear_128_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized, x = input_759_cast_fp16)[name = string("linear_128_cast_fp16")];
+            fp16 var_3436_to_fp16 = const()[name = string("op_3436_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_3437_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3436_to_fp16)[name = string("op_3437_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3437_cast_fp16)[name = string("input_765_cast_fp16")];
+            tensor<int32, [1]> key_29_axes_0 = const()[name = string("key_29_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292501504)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292503616)))];
+            tensor<fp16, [1, 56, 1024]> key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = string("key_29_cast_fp16")];
+            bool input_767_interleave_0 = const()[name = string("input_767_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_767_cast_fp16 = concat(axis = var_67, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = string("input_767_cast_fp16")];
+            bool var_3465_interleave_0 = const()[name = string("op_3465_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_3465_cast_fp16 = concat(axis = var_67, interleave = var_3465_interleave_0, values = key_29_cast_fp16)[name = string("op_3465_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292505728))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293292224))))[name = string("encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293292416)))];
+            tensor<fp16, [1, 56, 1024]> linear_129_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized, x = key_29_cast_fp16)[name = string("linear_129_cast_fp16")];
+            tensor<int32, [4]> var_3470 = const()[name = string("op_3470"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_85_cast_fp16 = reshape(shape = var_3470, x = linear_129_cast_fp16)[name = string("q_85_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293294528))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294081024))))[name = string("encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294081216)))];
+            tensor<fp16, [1, 98, 1024]> linear_130_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_130_cast_fp16")];
+            tensor<int32, [4]> var_3475 = const()[name = string("op_3475"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_57_cast_fp16 = reshape(shape = var_3475, x = linear_130_cast_fp16)[name = string("k_57_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294083328))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294869824))))[name = string("encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294870016)))];
+            tensor<fp16, [1, 98, 1024]> linear_131_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_131_cast_fp16")];
+            tensor<int32, [4]> var_3480 = const()[name = string("op_3480"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_29_cast_fp16 = reshape(shape = var_3480, x = linear_131_cast_fp16)[name = string("v_29_cast_fp16")];
+            tensor<int32, [4]> value_37_perm_0 = const()[name = string("value_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294872128)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3493_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = string("op_3493_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294874240)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3495_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = string("op_3495_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_29_perm_0 = const()[name = string("q_with_bias_v_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_371_transpose_x_0 = const()[name = string("x_371_transpose_x_0"), val = bool(false)];
+            bool x_371_transpose_y_0 = const()[name = string("x_371_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_3497_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294876352))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295076096))))[name = string("op_3497_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3495_cast_fp16)[name = string("transpose_236")];
+            tensor<fp16, [1, 8, 56, 195]> x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = op_3497_to_fp16_quantized)[name = string("x_371_cast_fp16")];
+            tensor<int32, [8]> x_373_pad_0 = const()[name = string("x_373_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_373_mode_0 = const()[name = string("x_373_mode_0"), val = string("constant")];
+            fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_373_cast_fp16 = pad(constant_val = const_261_to_fp16, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371_cast_fp16)[name = string("x_373_cast_fp16")];
+            tensor<int32, [4]> var_3505 = const()[name = string("op_3505"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_375_cast_fp16 = reshape(shape = var_3505, x = x_373_cast_fp16)[name = string("x_375_cast_fp16")];
+            tensor<int32, [4]> var_3509_begin_0 = const()[name = string("op_3509_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3509_end_0 = const()[name = string("op_3509_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_3509_end_mask_0 = const()[name = string("op_3509_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_3509_cast_fp16 = slice_by_index(begin = var_3509_begin_0, end = var_3509_end_0, end_mask = var_3509_end_mask_0, x = x_375_cast_fp16)[name = string("op_3509_cast_fp16")];
+            tensor<int32, [4]> var_3510 = const()[name = string("op_3510"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_57_cast_fp16 = reshape(shape = var_3510, x = var_3509_cast_fp16)[name = string("matrix_bd_57_cast_fp16")];
+            bool matrix_ac_29_transpose_x_0 = const()[name = string("matrix_ac_29_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_29_transpose_y_0 = const()[name = string("matrix_ac_29_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_124_perm_0 = const()[name = string("transpose_124_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_125_perm_0 = const()[name = string("transpose_125_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_125 = transpose(perm = transpose_125_perm_0, x = k_57_cast_fp16)[name = string("transpose_234")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_124 = transpose(perm = transpose_124_perm_0, x = var_3493_cast_fp16)[name = string("transpose_235")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_29_cast_fp16 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_124, y = transpose_125)[name = string("matrix_ac_29_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_59_begin_0 = const()[name = string("matrix_bd_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_59_end_0 = const()[name = string("matrix_bd_59_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_59_end_mask_0 = const()[name = string("matrix_bd_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_59_cast_fp16 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57_cast_fp16)[name = string("matrix_bd_59_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3519_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = string("op_3519_cast_fp16")];
+            fp16 _inversed_scores_57_y_0_to_fp16 = const()[name = string("_inversed_scores_57_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_57_cast_fp16 = mul(x = var_3519_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = string("_inversed_scores_57_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_59_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_11)[name = string("scores_59_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3525_cast_fp16 = softmax(axis = var_58, x = scores_59_cast_fp16)[name = string("op_3525_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_769_cast_fp16 = select(a = var_43_to_fp16, b = var_3525_cast_fp16, cond = mask_11)[name = string("input_769_cast_fp16")];
+            bool x_377_transpose_x_0 = const()[name = string("x_377_transpose_x_0"), val = bool(false)];
+            bool x_377_transpose_y_0 = const()[name = string("x_377_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = v_29_cast_fp16)[name = string("transpose_233")];
+            tensor<fp16, [1, 8, 56, 128]> x_377_cast_fp16 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769_cast_fp16, y = value_37_cast_fp16)[name = string("x_377_cast_fp16")];
+            tensor<int32, [4]> var_3529_perm_0 = const()[name = string("op_3529_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3530 = const()[name = string("op_3530"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_3529_cast_fp16 = transpose(perm = var_3529_perm_0, x = x_377_cast_fp16)[name = string("transpose_232")];
+            tensor<fp16, [1, 56, 1024]> input_771_cast_fp16 = reshape(shape = var_3530, x = var_3529_cast_fp16)[name = string("input_771_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295076608))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295863104))))[name = string("encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295863296)))];
+            tensor<fp16, [1, 56, 1024]> linear_133_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized, x = input_771_cast_fp16)[name = string("linear_133_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = string("input_775_cast_fp16")];
+            tensor<int32, [1]> x_381_axes_0 = const()[name = string("x_381_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295865408)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295867520)))];
+            tensor<fp16, [1, 56, 1024]> x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = string("x_381_cast_fp16")];
+            tensor<int32, [3]> input_777_perm_0 = const()[name = string("input_777_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_779_pad_type_0 = const()[name = string("input_779_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_779_strides_0 = const()[name = string("input_779_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_779_pad_0 = const()[name = string("input_779_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_779_dilations_0 = const()[name = string("input_779_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_779_groups_0 = const()[name = string("input_779_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295869632))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297966848))))[name = string("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = string("transpose_231")];
+            tensor<fp16, [1, 2048, 56]> input_779_cast_fp16 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_777_cast_fp16)[name = string("input_779_cast_fp16")];
+            int32 x_383_split_num_splits_0 = const()[name = string("x_383_split_num_splits_0"), val = int32(2)];
+            int32 x_383_split_axis_0 = const()[name = string("x_383_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_383_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_383_split_cast_fp16_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779_cast_fp16)[name = string("x_383_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = string("x_383_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = string("x_383_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_781_cast_fp16 = select(a = var_43_to_fp16, b = x_383_cast_fp16, cond = var_574)[name = string("input_781_cast_fp16")];
+            bool new_x_59_interleave_0 = const()[name = string("new_x_59_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_59_cast_fp16 = concat(axis = var_58, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = string("new_x_59_cast_fp16")];
+            tensor<int32, [3]> var_3569_begin_0 = const()[name = string("op_3569_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_3569_end_0 = const()[name = string("op_3569_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_3569_end_mask_0 = const()[name = string("op_3569_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3569_cast_fp16 = slice_by_index(begin = var_3569_begin_0, end = var_3569_end_0, end_mask = var_3569_end_mask_0, x = new_x_59_cast_fp16)[name = string("op_3569_cast_fp16")];
+            string x_385_pad_type_0 = const()[name = string("x_385_pad_type_0"), val = string("valid")];
+            int32 x_385_groups_0 = const()[name = string("x_385_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_385_strides_0 = const()[name = string("x_385_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_385_pad_0 = const()[name = string("x_385_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_385_dilations_0 = const()[name = string("x_385_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297971008))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297980288))))[name = string("encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_59_cast_fp16)[name = string("x_385_cast_fp16")];
+            tensor<int32, [3]> input_783_perm_0 = const()[name = string("input_783_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_387_axes_0 = const()[name = string("x_387_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297982400)))];
+            tensor<fp16, [1024]> encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297984512)))];
+            tensor<fp16, [1, 56, 1024]> input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = string("transpose_230")];
+            tensor<fp16, [1, 56, 1024]> x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = string("x_387_cast_fp16")];
+            tensor<int32, [3]> input_785_perm_0 = const()[name = string("input_785_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = string("transpose_229")];
+            tensor<fp16, [1, 1024, 56]> input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = string("input_787_cast_fp16")];
+            string x_389_pad_type_0 = const()[name = string("x_389_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_389_strides_0 = const()[name = string("x_389_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_389_pad_0 = const()[name = string("x_389_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_389_dilations_0 = const()[name = string("x_389_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_389_groups_0 = const()[name = string("x_389_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297986624))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299035264))))[name = string("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_389_cast_fp16 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_787_cast_fp16)[name = string("x_389_cast_fp16")];
+            tensor<int32, [3]> input_789_perm_0 = const()[name = string("input_789_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = string("transpose_228")];
+            tensor<fp16, [1, 56, 1024]> input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = string("input_791_cast_fp16")];
+            tensor<int32, [1]> input_793_axes_0 = const()[name = string("input_793_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299037376)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299039488)))];
+            tensor<fp16, [1, 56, 1024]> input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = string("input_793_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299041600))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302187392))))[name = string("encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_14_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302187584)))];
+            tensor<fp16, [1, 56, 4096]> linear_134_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized, x = input_793_cast_fp16)[name = string("linear_134_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = string("input_797_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302195840))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305341632))))[name = string("encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305341824)))];
+            tensor<fp16, [1, 56, 1024]> linear_135_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized, x = input_797_cast_fp16)[name = string("linear_135_cast_fp16")];
+            fp16 var_3612_to_fp16 = const()[name = string("op_3612_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_3613_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3612_to_fp16)[name = string("op_3613_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3613_cast_fp16)[name = string("input_803_cast_fp16")];
+            tensor<int32, [1]> input_805_axes_0 = const()[name = string("input_805_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305343936)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305346048)))];
+            tensor<fp16, [1, 56, 1024]> input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = string("input_805_cast_fp16")];
+            tensor<int32, [4]> cache_61_begin_0 = const()[name = string("cache_61_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_61_end_0 = const()[name = string("cache_61_end_0"), val = tensor<int32, [4]>([16, 1, 42, 1024])];
+            tensor<bool, [4]> cache_61_end_mask_0 = const()[name = string("cache_61_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_61_squeeze_mask_0 = const()[name = string("cache_61_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_61_cast_fp16 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_61_cast_fp16")];
+            tensor<int32, [4]> cache_63_begin_0 = const()[name = string("cache_63_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_63_end_0 = const()[name = string("cache_63_end_0"), val = tensor<int32, [4]>([16, 1, 1024, 8])];
+            tensor<bool, [4]> cache_63_end_mask_0 = const()[name = string("cache_63_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_63_squeeze_mask_0 = const()[name = string("cache_63_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_63_cast_fp16 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_63_cast_fp16")];
+            tensor<int32, [1]> input_807_axes_0 = const()[name = string("input_807_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305348160)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305350272)))];
+            tensor<fp16, [1, 56, 1024]> input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = string("input_807_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(305352384))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308498176))))[name = string("encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_15_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308498368)))];
+            tensor<fp16, [1, 56, 4096]> linear_136_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized, x = input_807_cast_fp16)[name = string("linear_136_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = string("input_811_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308506624))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311652416))))[name = string("encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311652608)))];
+            tensor<fp16, [1, 56, 1024]> linear_137_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("linear_137_cast_fp16")];
+            fp16 var_3649_to_fp16 = const()[name = string("op_3649_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_3650_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3649_to_fp16)[name = string("op_3650_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3650_cast_fp16)[name = string("input_817_cast_fp16")];
+            tensor<int32, [1]> key_31_axes_0 = const()[name = string("key_31_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311654720)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311656832)))];
+            tensor<fp16, [1, 56, 1024]> key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = string("key_31_cast_fp16")];
+            bool input_819_interleave_0 = const()[name = string("input_819_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_819_cast_fp16 = concat(axis = var_67, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = string("input_819_cast_fp16")];
+            bool var_3678_interleave_0 = const()[name = string("op_3678_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_3678_cast_fp16 = concat(axis = var_67, interleave = var_3678_interleave_0, values = key_31_cast_fp16)[name = string("op_3678_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311658944))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312445440))))[name = string("encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312445632)))];
+            tensor<fp16, [1, 56, 1024]> linear_138_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized, x = key_31_cast_fp16)[name = string("linear_138_cast_fp16")];
+            tensor<int32, [4]> var_3683 = const()[name = string("op_3683"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_91_cast_fp16 = reshape(shape = var_3683, x = linear_138_cast_fp16)[name = string("q_91_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312447744))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313234240))))[name = string("encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313234432)))];
+            tensor<fp16, [1, 98, 1024]> linear_139_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_139_cast_fp16")];
+            tensor<int32, [4]> var_3688 = const()[name = string("op_3688"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_61_cast_fp16 = reshape(shape = var_3688, x = linear_139_cast_fp16)[name = string("k_61_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313236544))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314023040))))[name = string("encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314023232)))];
+            tensor<fp16, [1, 98, 1024]> linear_140_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_140_cast_fp16")];
+            tensor<int32, [4]> var_3693 = const()[name = string("op_3693"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_31_cast_fp16 = reshape(shape = var_3693, x = linear_140_cast_fp16)[name = string("v_31_cast_fp16")];
+            tensor<int32, [4]> value_39_perm_0 = const()[name = string("value_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314025344)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3706_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = string("op_3706_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314027456)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3708_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = string("op_3708_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_31_perm_0 = const()[name = string("q_with_bias_v_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_397_transpose_x_0 = const()[name = string("x_397_transpose_x_0"), val = bool(false)];
+            bool x_397_transpose_y_0 = const()[name = string("x_397_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_3710_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314029568))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314229312))))[name = string("op_3710_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3708_cast_fp16)[name = string("transpose_227")];
+            tensor<fp16, [1, 8, 56, 195]> x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = op_3710_to_fp16_quantized)[name = string("x_397_cast_fp16")];
+            tensor<int32, [8]> x_399_pad_0 = const()[name = string("x_399_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_399_mode_0 = const()[name = string("x_399_mode_0"), val = string("constant")];
+            fp16 const_274_to_fp16 = const()[name = string("const_274_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_399_cast_fp16 = pad(constant_val = const_274_to_fp16, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397_cast_fp16)[name = string("x_399_cast_fp16")];
+            tensor<int32, [4]> var_3718 = const()[name = string("op_3718"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_401_cast_fp16 = reshape(shape = var_3718, x = x_399_cast_fp16)[name = string("x_401_cast_fp16")];
+            tensor<int32, [4]> var_3722_begin_0 = const()[name = string("op_3722_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3722_end_0 = const()[name = string("op_3722_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_3722_end_mask_0 = const()[name = string("op_3722_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_3722_cast_fp16 = slice_by_index(begin = var_3722_begin_0, end = var_3722_end_0, end_mask = var_3722_end_mask_0, x = x_401_cast_fp16)[name = string("op_3722_cast_fp16")];
+            tensor<int32, [4]> var_3723 = const()[name = string("op_3723"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_61_cast_fp16 = reshape(shape = var_3723, x = var_3722_cast_fp16)[name = string("matrix_bd_61_cast_fp16")];
+            bool matrix_ac_31_transpose_x_0 = const()[name = string("matrix_ac_31_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_31_transpose_y_0 = const()[name = string("matrix_ac_31_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_126_perm_0 = const()[name = string("transpose_126_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_127_perm_0 = const()[name = string("transpose_127_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_127 = transpose(perm = transpose_127_perm_0, x = k_61_cast_fp16)[name = string("transpose_225")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_126 = transpose(perm = transpose_126_perm_0, x = var_3706_cast_fp16)[name = string("transpose_226")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_31_cast_fp16 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_126, y = transpose_127)[name = string("matrix_ac_31_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_63_begin_0 = const()[name = string("matrix_bd_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_63_end_0 = const()[name = string("matrix_bd_63_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_63_end_mask_0 = const()[name = string("matrix_bd_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_63_cast_fp16 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61_cast_fp16)[name = string("matrix_bd_63_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3732_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = string("op_3732_cast_fp16")];
+            fp16 _inversed_scores_61_y_0_to_fp16 = const()[name = string("_inversed_scores_61_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_61_cast_fp16 = mul(x = var_3732_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = string("_inversed_scores_61_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_63_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_11)[name = string("scores_63_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3738_cast_fp16 = softmax(axis = var_58, x = scores_63_cast_fp16)[name = string("op_3738_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_821_cast_fp16 = select(a = var_43_to_fp16, b = var_3738_cast_fp16, cond = mask_11)[name = string("input_821_cast_fp16")];
+            bool x_403_transpose_x_0 = const()[name = string("x_403_transpose_x_0"), val = bool(false)];
+            bool x_403_transpose_y_0 = const()[name = string("x_403_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = v_31_cast_fp16)[name = string("transpose_224")];
+            tensor<fp16, [1, 8, 56, 128]> x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821_cast_fp16, y = value_39_cast_fp16)[name = string("x_403_cast_fp16")];
+            tensor<int32, [4]> var_3742_perm_0 = const()[name = string("op_3742_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3743 = const()[name = string("op_3743"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_3742_cast_fp16 = transpose(perm = var_3742_perm_0, x = x_403_cast_fp16)[name = string("transpose_223")];
+            tensor<fp16, [1, 56, 1024]> input_823_cast_fp16 = reshape(shape = var_3743, x = var_3742_cast_fp16)[name = string("input_823_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314229824))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315016320))))[name = string("encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315016512)))];
+            tensor<fp16, [1, 56, 1024]> linear_142_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized, x = input_823_cast_fp16)[name = string("linear_142_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = string("input_827_cast_fp16")];
+            tensor<int32, [1]> x_407_axes_0 = const()[name = string("x_407_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315018624)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315020736)))];
+            tensor<fp16, [1, 56, 1024]> x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = string("x_407_cast_fp16")];
+            tensor<int32, [3]> input_829_perm_0 = const()[name = string("input_829_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_831_pad_type_0 = const()[name = string("input_831_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_831_strides_0 = const()[name = string("input_831_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_831_pad_0 = const()[name = string("input_831_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_831_dilations_0 = const()[name = string("input_831_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_831_groups_0 = const()[name = string("input_831_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315022848))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317120064))))[name = string("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = string("transpose_222")];
+            tensor<fp16, [1, 2048, 56]> input_831_cast_fp16 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_829_cast_fp16)[name = string("input_831_cast_fp16")];
+            int32 x_409_split_num_splits_0 = const()[name = string("x_409_split_num_splits_0"), val = int32(2)];
+            int32 x_409_split_axis_0 = const()[name = string("x_409_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_409_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_409_split_cast_fp16_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831_cast_fp16)[name = string("x_409_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = string("x_409_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = string("x_409_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_833_cast_fp16 = select(a = var_43_to_fp16, b = x_409_cast_fp16, cond = var_574)[name = string("input_833_cast_fp16")];
+            bool new_x_63_interleave_0 = const()[name = string("new_x_63_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_63_cast_fp16 = concat(axis = var_58, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = string("new_x_63_cast_fp16")];
+            tensor<int32, [3]> var_3782_begin_0 = const()[name = string("op_3782_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_3782_end_0 = const()[name = string("op_3782_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_3782_end_mask_0 = const()[name = string("op_3782_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3782_cast_fp16 = slice_by_index(begin = var_3782_begin_0, end = var_3782_end_0, end_mask = var_3782_end_mask_0, x = new_x_63_cast_fp16)[name = string("op_3782_cast_fp16")];
+            string x_411_pad_type_0 = const()[name = string("x_411_pad_type_0"), val = string("valid")];
+            int32 x_411_groups_0 = const()[name = string("x_411_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_411_strides_0 = const()[name = string("x_411_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_411_pad_0 = const()[name = string("x_411_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_411_dilations_0 = const()[name = string("x_411_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317124224))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317133504))))[name = string("encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_411_cast_fp16 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_63_cast_fp16)[name = string("x_411_cast_fp16")];
+            tensor<int32, [3]> input_835_perm_0 = const()[name = string("input_835_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_413_axes_0 = const()[name = string("x_413_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317135616)))];
+            tensor<fp16, [1024]> encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317137728)))];
+            tensor<fp16, [1, 56, 1024]> input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = string("transpose_221")];
+            tensor<fp16, [1, 56, 1024]> x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = string("x_413_cast_fp16")];
+            tensor<int32, [3]> input_837_perm_0 = const()[name = string("input_837_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = string("transpose_220")];
+            tensor<fp16, [1, 1024, 56]> input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = string("input_839_cast_fp16")];
+            string x_415_pad_type_0 = const()[name = string("x_415_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_415_strides_0 = const()[name = string("x_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_415_pad_0 = const()[name = string("x_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_415_dilations_0 = const()[name = string("x_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_415_groups_0 = const()[name = string("x_415_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317139840))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318188480))))[name = string("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_839_cast_fp16)[name = string("x_415_cast_fp16")];
+            tensor<int32, [3]> input_841_perm_0 = const()[name = string("input_841_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = string("transpose_219")];
+            tensor<fp16, [1, 56, 1024]> input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = string("input_843_cast_fp16")];
+            tensor<int32, [1]> input_845_axes_0 = const()[name = string("input_845_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318190592)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318192704)))];
+            tensor<fp16, [1, 56, 1024]> input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = string("input_845_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318194816))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321340608))))[name = string("encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_15_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321340800)))];
+            tensor<fp16, [1, 56, 4096]> linear_143_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized, x = input_845_cast_fp16)[name = string("linear_143_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = string("input_849_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321349056))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324494848))))[name = string("encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324495040)))];
+            tensor<fp16, [1, 56, 1024]> linear_144_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized, x = input_849_cast_fp16)[name = string("linear_144_cast_fp16")];
+            fp16 var_3825_to_fp16 = const()[name = string("op_3825_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_3826_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3825_to_fp16)[name = string("op_3826_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3826_cast_fp16)[name = string("input_855_cast_fp16")];
+            tensor<int32, [1]> input_857_axes_0 = const()[name = string("input_857_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324497152)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324499264)))];
+            tensor<fp16, [1, 56, 1024]> input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = string("input_857_cast_fp16")];
+            tensor<int32, [4]> cache_65_begin_0 = const()[name = string("cache_65_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_65_end_0 = const()[name = string("cache_65_end_0"), val = tensor<int32, [4]>([17, 1, 42, 1024])];
+            tensor<bool, [4]> cache_65_end_mask_0 = const()[name = string("cache_65_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_65_squeeze_mask_0 = const()[name = string("cache_65_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_65_cast_fp16 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_65_cast_fp16")];
+            tensor<int32, [4]> cache_67_begin_0 = const()[name = string("cache_67_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_67_end_0 = const()[name = string("cache_67_end_0"), val = tensor<int32, [4]>([17, 1, 1024, 8])];
+            tensor<bool, [4]> cache_67_end_mask_0 = const()[name = string("cache_67_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_67_squeeze_mask_0 = const()[name = string("cache_67_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_67_cast_fp16 = slice_by_index(begin = cache_67_begin_0, end = cache_67_end_0, end_mask = cache_67_end_mask_0, squeeze_mask = cache_67_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_67_cast_fp16")];
+            tensor<int32, [1]> input_859_axes_0 = const()[name = string("input_859_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324501376)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324503488)))];
+            tensor<fp16, [1, 56, 1024]> input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = string("input_859_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324505600))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327651392))))[name = string("encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_16_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327651584)))];
+            tensor<fp16, [1, 56, 4096]> linear_145_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized, x = input_859_cast_fp16)[name = string("linear_145_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = string("input_863_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327659840))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330805632))))[name = string("encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330805824)))];
+            tensor<fp16, [1, 56, 1024]> linear_146_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized, x = input_863_cast_fp16)[name = string("linear_146_cast_fp16")];
+            fp16 var_3862_to_fp16 = const()[name = string("op_3862_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_3863_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3862_to_fp16)[name = string("op_3863_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3863_cast_fp16)[name = string("input_869_cast_fp16")];
+            tensor<int32, [1]> key_33_axes_0 = const()[name = string("key_33_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330807936)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330810048)))];
+            tensor<fp16, [1, 56, 1024]> key_33_cast_fp16 = layer_norm(axes = key_33_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = string("key_33_cast_fp16")];
+            bool input_871_interleave_0 = const()[name = string("input_871_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_871_cast_fp16 = concat(axis = var_67, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_33_cast_fp16))[name = string("input_871_cast_fp16")];
+            bool var_3891_interleave_0 = const()[name = string("op_3891_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_3891_cast_fp16 = concat(axis = var_67, interleave = var_3891_interleave_0, values = key_33_cast_fp16)[name = string("op_3891_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330812160))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331598656))))[name = string("encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331598848)))];
+            tensor<fp16, [1, 56, 1024]> linear_147_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized, x = key_33_cast_fp16)[name = string("linear_147_cast_fp16")];
+            tensor<int32, [4]> var_3896 = const()[name = string("op_3896"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_97_cast_fp16 = reshape(shape = var_3896, x = linear_147_cast_fp16)[name = string("q_97_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331600960))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332387456))))[name = string("encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332387648)))];
+            tensor<fp16, [1, 98, 1024]> linear_148_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_148_cast_fp16")];
+            tensor<int32, [4]> var_3901 = const()[name = string("op_3901"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_65_cast_fp16 = reshape(shape = var_3901, x = linear_148_cast_fp16)[name = string("k_65_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332389760))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333176256))))[name = string("encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333176448)))];
+            tensor<fp16, [1, 98, 1024]> linear_149_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_149_cast_fp16")];
+            tensor<int32, [4]> var_3906 = const()[name = string("op_3906"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_33_cast_fp16 = reshape(shape = var_3906, x = linear_149_cast_fp16)[name = string("v_33_cast_fp16")];
+            tensor<int32, [4]> value_41_perm_0 = const()[name = string("value_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333178560)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3919_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = string("op_3919_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333180672)))];
+            tensor<fp16, [1, 56, 8, 128]> var_3921_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = string("op_3921_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_33_perm_0 = const()[name = string("q_with_bias_v_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_423_transpose_x_0 = const()[name = string("x_423_transpose_x_0"), val = bool(false)];
+            bool x_423_transpose_y_0 = const()[name = string("x_423_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_3923_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333182784))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333382528))))[name = string("op_3923_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_33_cast_fp16 = transpose(perm = q_with_bias_v_33_perm_0, x = var_3921_cast_fp16)[name = string("transpose_218")];
+            tensor<fp16, [1, 8, 56, 195]> x_423_cast_fp16 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_33_cast_fp16, y = op_3923_to_fp16_quantized)[name = string("x_423_cast_fp16")];
+            tensor<int32, [8]> x_425_pad_0 = const()[name = string("x_425_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_425_mode_0 = const()[name = string("x_425_mode_0"), val = string("constant")];
+            fp16 const_287_to_fp16 = const()[name = string("const_287_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_425_cast_fp16 = pad(constant_val = const_287_to_fp16, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423_cast_fp16)[name = string("x_425_cast_fp16")];
+            tensor<int32, [4]> var_3931 = const()[name = string("op_3931"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_427_cast_fp16 = reshape(shape = var_3931, x = x_425_cast_fp16)[name = string("x_427_cast_fp16")];
+            tensor<int32, [4]> var_3935_begin_0 = const()[name = string("op_3935_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3935_end_0 = const()[name = string("op_3935_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_3935_end_mask_0 = const()[name = string("op_3935_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_3935_cast_fp16 = slice_by_index(begin = var_3935_begin_0, end = var_3935_end_0, end_mask = var_3935_end_mask_0, x = x_427_cast_fp16)[name = string("op_3935_cast_fp16")];
+            tensor<int32, [4]> var_3936 = const()[name = string("op_3936"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_65_cast_fp16 = reshape(shape = var_3936, x = var_3935_cast_fp16)[name = string("matrix_bd_65_cast_fp16")];
+            bool matrix_ac_33_transpose_x_0 = const()[name = string("matrix_ac_33_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_33_transpose_y_0 = const()[name = string("matrix_ac_33_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_128_perm_0 = const()[name = string("transpose_128_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_129_perm_0 = const()[name = string("transpose_129_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_129 = transpose(perm = transpose_129_perm_0, x = k_65_cast_fp16)[name = string("transpose_216")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_128 = transpose(perm = transpose_128_perm_0, x = var_3919_cast_fp16)[name = string("transpose_217")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_33_cast_fp16 = matmul(transpose_x = matrix_ac_33_transpose_x_0, transpose_y = matrix_ac_33_transpose_y_0, x = transpose_128, y = transpose_129)[name = string("matrix_ac_33_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_67_begin_0 = const()[name = string("matrix_bd_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_67_end_0 = const()[name = string("matrix_bd_67_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_67_end_mask_0 = const()[name = string("matrix_bd_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_67_cast_fp16 = slice_by_index(begin = matrix_bd_67_begin_0, end = matrix_bd_67_end_0, end_mask = matrix_bd_67_end_mask_0, x = matrix_bd_65_cast_fp16)[name = string("matrix_bd_67_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3945_cast_fp16 = add(x = matrix_ac_33_cast_fp16, y = matrix_bd_67_cast_fp16)[name = string("op_3945_cast_fp16")];
+            fp16 _inversed_scores_65_y_0_to_fp16 = const()[name = string("_inversed_scores_65_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_65_cast_fp16 = mul(x = var_3945_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = string("_inversed_scores_65_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_67_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_11)[name = string("scores_67_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_3951_cast_fp16 = softmax(axis = var_58, x = scores_67_cast_fp16)[name = string("op_3951_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_873_cast_fp16 = select(a = var_43_to_fp16, b = var_3951_cast_fp16, cond = mask_11)[name = string("input_873_cast_fp16")];
+            bool x_429_transpose_x_0 = const()[name = string("x_429_transpose_x_0"), val = bool(false)];
+            bool x_429_transpose_y_0 = const()[name = string("x_429_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = v_33_cast_fp16)[name = string("transpose_215")];
+            tensor<fp16, [1, 8, 56, 128]> x_429_cast_fp16 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873_cast_fp16, y = value_41_cast_fp16)[name = string("x_429_cast_fp16")];
+            tensor<int32, [4]> var_3955_perm_0 = const()[name = string("op_3955_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3956 = const()[name = string("op_3956"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_3955_cast_fp16 = transpose(perm = var_3955_perm_0, x = x_429_cast_fp16)[name = string("transpose_214")];
+            tensor<fp16, [1, 56, 1024]> input_875_cast_fp16 = reshape(shape = var_3956, x = var_3955_cast_fp16)[name = string("input_875_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333383040))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334169536))))[name = string("encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334169728)))];
+            tensor<fp16, [1, 56, 1024]> linear_151_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized, x = input_875_cast_fp16)[name = string("linear_151_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = string("input_879_cast_fp16")];
+            tensor<int32, [1]> x_433_axes_0 = const()[name = string("x_433_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334171840)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334173952)))];
+            tensor<fp16, [1, 56, 1024]> x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = string("x_433_cast_fp16")];
+            tensor<int32, [3]> input_881_perm_0 = const()[name = string("input_881_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_883_pad_type_0 = const()[name = string("input_883_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_883_strides_0 = const()[name = string("input_883_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_883_pad_0 = const()[name = string("input_883_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_883_dilations_0 = const()[name = string("input_883_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_883_groups_0 = const()[name = string("input_883_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334176064))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336273280))))[name = string("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = string("transpose_213")];
+            tensor<fp16, [1, 2048, 56]> input_883_cast_fp16 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_881_cast_fp16)[name = string("input_883_cast_fp16")];
+            int32 x_435_split_num_splits_0 = const()[name = string("x_435_split_num_splits_0"), val = int32(2)];
+            int32 x_435_split_axis_0 = const()[name = string("x_435_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_435_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_435_split_cast_fp16_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883_cast_fp16)[name = string("x_435_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = string("x_435_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = string("x_435_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_885_cast_fp16 = select(a = var_43_to_fp16, b = x_435_cast_fp16, cond = var_574)[name = string("input_885_cast_fp16")];
+            bool new_x_67_interleave_0 = const()[name = string("new_x_67_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_67_cast_fp16 = concat(axis = var_58, interleave = new_x_67_interleave_0, values = (cache_67_cast_fp16, input_885_cast_fp16))[name = string("new_x_67_cast_fp16")];
+            tensor<int32, [3]> var_3995_begin_0 = const()[name = string("op_3995_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_3995_end_0 = const()[name = string("op_3995_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_3995_end_mask_0 = const()[name = string("op_3995_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3995_cast_fp16 = slice_by_index(begin = var_3995_begin_0, end = var_3995_end_0, end_mask = var_3995_end_mask_0, x = new_x_67_cast_fp16)[name = string("op_3995_cast_fp16")];
+            string x_437_pad_type_0 = const()[name = string("x_437_pad_type_0"), val = string("valid")];
+            int32 x_437_groups_0 = const()[name = string("x_437_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_437_strides_0 = const()[name = string("x_437_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_437_pad_0 = const()[name = string("x_437_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_437_dilations_0 = const()[name = string("x_437_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336277440))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336286720))))[name = string("encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_437_cast_fp16 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_67_cast_fp16)[name = string("x_437_cast_fp16")];
+            tensor<int32, [3]> input_887_perm_0 = const()[name = string("input_887_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_439_axes_0 = const()[name = string("x_439_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336288832)))];
+            tensor<fp16, [1024]> encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336290944)))];
+            tensor<fp16, [1, 56, 1024]> input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = string("transpose_212")];
+            tensor<fp16, [1, 56, 1024]> x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = string("x_439_cast_fp16")];
+            tensor<int32, [3]> input_889_perm_0 = const()[name = string("input_889_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = string("transpose_211")];
+            tensor<fp16, [1, 1024, 56]> input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = string("input_891_cast_fp16")];
+            string x_441_pad_type_0 = const()[name = string("x_441_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_441_strides_0 = const()[name = string("x_441_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_441_pad_0 = const()[name = string("x_441_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_441_dilations_0 = const()[name = string("x_441_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_441_groups_0 = const()[name = string("x_441_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336293056))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337341696))))[name = string("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_441_cast_fp16 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_891_cast_fp16)[name = string("x_441_cast_fp16")];
+            tensor<int32, [3]> input_893_perm_0 = const()[name = string("input_893_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = string("transpose_210")];
+            tensor<fp16, [1, 56, 1024]> input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = string("input_895_cast_fp16")];
+            tensor<int32, [1]> input_897_axes_0 = const()[name = string("input_897_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337343808)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337345920)))];
+            tensor<fp16, [1, 56, 1024]> input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = string("input_897_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(337348032))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340493824))))[name = string("encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_16_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340494016)))];
+            tensor<fp16, [1, 56, 4096]> linear_152_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized, x = input_897_cast_fp16)[name = string("linear_152_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = string("input_901_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(340502272))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343648064))))[name = string("encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343648256)))];
+            tensor<fp16, [1, 56, 1024]> linear_153_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized, x = input_901_cast_fp16)[name = string("linear_153_cast_fp16")];
+            fp16 var_4038_to_fp16 = const()[name = string("op_4038_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4039_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_4038_to_fp16)[name = string("op_4039_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_907_cast_fp16 = add(x = input_895_cast_fp16, y = var_4039_cast_fp16)[name = string("input_907_cast_fp16")];
+            tensor<int32, [1]> input_909_axes_0 = const()[name = string("input_909_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343650368)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343652480)))];
+            tensor<fp16, [1, 56, 1024]> input_909_cast_fp16 = layer_norm(axes = input_909_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_907_cast_fp16)[name = string("input_909_cast_fp16")];
+            tensor<int32, [4]> cache_69_begin_0 = const()[name = string("cache_69_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> cache_69_end_0 = const()[name = string("cache_69_end_0"), val = tensor<int32, [4]>([18, 1, 42, 1024])];
+            tensor<bool, [4]> cache_69_end_mask_0 = const()[name = string("cache_69_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_69_squeeze_mask_0 = const()[name = string("cache_69_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_69_cast_fp16 = slice_by_index(begin = cache_69_begin_0, end = cache_69_end_0, end_mask = cache_69_end_mask_0, squeeze_mask = cache_69_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_69_cast_fp16")];
+            tensor<int32, [4]> cache_71_begin_0 = const()[name = string("cache_71_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> cache_71_end_0 = const()[name = string("cache_71_end_0"), val = tensor<int32, [4]>([18, 1, 1024, 8])];
+            tensor<bool, [4]> cache_71_end_mask_0 = const()[name = string("cache_71_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_71_squeeze_mask_0 = const()[name = string("cache_71_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_71_cast_fp16 = slice_by_index(begin = cache_71_begin_0, end = cache_71_end_0, end_mask = cache_71_end_mask_0, squeeze_mask = cache_71_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_71_cast_fp16")];
+            tensor<int32, [1]> input_911_axes_0 = const()[name = string("input_911_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343654592)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343656704)))];
+            tensor<fp16, [1, 56, 1024]> input_911_cast_fp16 = layer_norm(axes = input_911_axes_0, beta = encoder_layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_17_norm_feed_forward1_weight_to_fp16, x = input_909_cast_fp16)[name = string("input_911_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(343658816))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346804608))))[name = string("encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_17_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346804800)))];
+            tensor<fp16, [1, 56, 4096]> linear_154_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized, x = input_911_cast_fp16)[name = string("linear_154_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_915_cast_fp16 = silu(x = linear_154_cast_fp16)[name = string("input_915_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346813056))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349958848))))[name = string("encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349959040)))];
+            tensor<fp16, [1, 56, 1024]> linear_155_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized, x = input_915_cast_fp16)[name = string("linear_155_cast_fp16")];
+            fp16 var_4075_to_fp16 = const()[name = string("op_4075_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4076_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_4075_to_fp16)[name = string("op_4076_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_921_cast_fp16 = add(x = input_909_cast_fp16, y = var_4076_cast_fp16)[name = string("input_921_cast_fp16")];
+            tensor<int32, [1]> key_35_axes_0 = const()[name = string("key_35_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349961152)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349963264)))];
+            tensor<fp16, [1, 56, 1024]> key_35_cast_fp16 = layer_norm(axes = key_35_axes_0, beta = encoder_layers_17_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_17_norm_self_att_weight_to_fp16, x = input_921_cast_fp16)[name = string("key_35_cast_fp16")];
+            bool input_923_interleave_0 = const()[name = string("input_923_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_923_cast_fp16 = concat(axis = var_67, interleave = input_923_interleave_0, values = (cache_69_cast_fp16, key_35_cast_fp16))[name = string("input_923_cast_fp16")];
+            bool var_4104_interleave_0 = const()[name = string("op_4104_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_4104_cast_fp16 = concat(axis = var_67, interleave = var_4104_interleave_0, values = key_35_cast_fp16)[name = string("op_4104_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349965376))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350751872))))[name = string("encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350752064)))];
+            tensor<fp16, [1, 56, 1024]> linear_156_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized, x = key_35_cast_fp16)[name = string("linear_156_cast_fp16")];
+            tensor<int32, [4]> var_4109 = const()[name = string("op_4109"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_103_cast_fp16 = reshape(shape = var_4109, x = linear_156_cast_fp16)[name = string("q_103_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350754176))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351540672))))[name = string("encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351540864)))];
+            tensor<fp16, [1, 98, 1024]> linear_157_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_157_cast_fp16")];
+            tensor<int32, [4]> var_4114 = const()[name = string("op_4114"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_69_cast_fp16 = reshape(shape = var_4114, x = linear_157_cast_fp16)[name = string("k_69_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351542976))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352329472))))[name = string("encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352329664)))];
+            tensor<fp16, [1, 98, 1024]> linear_158_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_158_cast_fp16")];
+            tensor<int32, [4]> var_4119 = const()[name = string("op_4119"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_35_cast_fp16 = reshape(shape = var_4119, x = linear_158_cast_fp16)[name = string("v_35_cast_fp16")];
+            tensor<int32, [4]> value_43_perm_0 = const()[name = string("value_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_17_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352331776)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4132_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_u_to_fp16)[name = string("op_4132_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_17_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352333888)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4134_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_v_to_fp16)[name = string("op_4134_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_35_perm_0 = const()[name = string("q_with_bias_v_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_449_transpose_x_0 = const()[name = string("x_449_transpose_x_0"), val = bool(false)];
+            bool x_449_transpose_y_0 = const()[name = string("x_449_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_4136_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352336000))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352535744))))[name = string("op_4136_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_35_cast_fp16 = transpose(perm = q_with_bias_v_35_perm_0, x = var_4134_cast_fp16)[name = string("transpose_209")];
+            tensor<fp16, [1, 8, 56, 195]> x_449_cast_fp16 = matmul(transpose_x = x_449_transpose_x_0, transpose_y = x_449_transpose_y_0, x = q_with_bias_v_35_cast_fp16, y = op_4136_to_fp16_quantized)[name = string("x_449_cast_fp16")];
+            tensor<int32, [8]> x_451_pad_0 = const()[name = string("x_451_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_451_mode_0 = const()[name = string("x_451_mode_0"), val = string("constant")];
+            fp16 const_300_to_fp16 = const()[name = string("const_300_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_451_cast_fp16 = pad(constant_val = const_300_to_fp16, mode = x_451_mode_0, pad = x_451_pad_0, x = x_449_cast_fp16)[name = string("x_451_cast_fp16")];
+            tensor<int32, [4]> var_4144 = const()[name = string("op_4144"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_453_cast_fp16 = reshape(shape = var_4144, x = x_451_cast_fp16)[name = string("x_453_cast_fp16")];
+            tensor<int32, [4]> var_4148_begin_0 = const()[name = string("op_4148_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4148_end_0 = const()[name = string("op_4148_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_4148_end_mask_0 = const()[name = string("op_4148_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_4148_cast_fp16 = slice_by_index(begin = var_4148_begin_0, end = var_4148_end_0, end_mask = var_4148_end_mask_0, x = x_453_cast_fp16)[name = string("op_4148_cast_fp16")];
+            tensor<int32, [4]> var_4149 = const()[name = string("op_4149"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_69_cast_fp16 = reshape(shape = var_4149, x = var_4148_cast_fp16)[name = string("matrix_bd_69_cast_fp16")];
+            bool matrix_ac_35_transpose_x_0 = const()[name = string("matrix_ac_35_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_35_transpose_y_0 = const()[name = string("matrix_ac_35_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_130_perm_0 = const()[name = string("transpose_130_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_131_perm_0 = const()[name = string("transpose_131_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_131 = transpose(perm = transpose_131_perm_0, x = k_69_cast_fp16)[name = string("transpose_207")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_130 = transpose(perm = transpose_130_perm_0, x = var_4132_cast_fp16)[name = string("transpose_208")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_35_cast_fp16 = matmul(transpose_x = matrix_ac_35_transpose_x_0, transpose_y = matrix_ac_35_transpose_y_0, x = transpose_130, y = transpose_131)[name = string("matrix_ac_35_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_71_begin_0 = const()[name = string("matrix_bd_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_71_end_0 = const()[name = string("matrix_bd_71_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_71_end_mask_0 = const()[name = string("matrix_bd_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_71_cast_fp16 = slice_by_index(begin = matrix_bd_71_begin_0, end = matrix_bd_71_end_0, end_mask = matrix_bd_71_end_mask_0, x = matrix_bd_69_cast_fp16)[name = string("matrix_bd_71_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_4158_cast_fp16 = add(x = matrix_ac_35_cast_fp16, y = matrix_bd_71_cast_fp16)[name = string("op_4158_cast_fp16")];
+            fp16 _inversed_scores_69_y_0_to_fp16 = const()[name = string("_inversed_scores_69_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_69_cast_fp16 = mul(x = var_4158_cast_fp16, y = _inversed_scores_69_y_0_to_fp16)[name = string("_inversed_scores_69_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_71_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_69_cast_fp16, cond = mask_11)[name = string("scores_71_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_4164_cast_fp16 = softmax(axis = var_58, x = scores_71_cast_fp16)[name = string("op_4164_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_925_cast_fp16 = select(a = var_43_to_fp16, b = var_4164_cast_fp16, cond = mask_11)[name = string("input_925_cast_fp16")];
+            bool x_455_transpose_x_0 = const()[name = string("x_455_transpose_x_0"), val = bool(false)];
+            bool x_455_transpose_y_0 = const()[name = string("x_455_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = v_35_cast_fp16)[name = string("transpose_206")];
+            tensor<fp16, [1, 8, 56, 128]> x_455_cast_fp16 = matmul(transpose_x = x_455_transpose_x_0, transpose_y = x_455_transpose_y_0, x = input_925_cast_fp16, y = value_43_cast_fp16)[name = string("x_455_cast_fp16")];
+            tensor<int32, [4]> var_4168_perm_0 = const()[name = string("op_4168_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4169 = const()[name = string("op_4169"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_4168_cast_fp16 = transpose(perm = var_4168_perm_0, x = x_455_cast_fp16)[name = string("transpose_205")];
+            tensor<fp16, [1, 56, 1024]> input_927_cast_fp16 = reshape(shape = var_4169, x = var_4168_cast_fp16)[name = string("input_927_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352536256))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353322752))))[name = string("encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353322944)))];
+            tensor<fp16, [1, 56, 1024]> linear_160_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized, x = input_927_cast_fp16)[name = string("linear_160_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_931_cast_fp16 = add(x = input_921_cast_fp16, y = linear_160_cast_fp16)[name = string("input_931_cast_fp16")];
+            tensor<int32, [1]> x_459_axes_0 = const()[name = string("x_459_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353325056)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353327168)))];
+            tensor<fp16, [1, 56, 1024]> x_459_cast_fp16 = layer_norm(axes = x_459_axes_0, beta = encoder_layers_17_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_17_norm_conv_weight_to_fp16, x = input_931_cast_fp16)[name = string("x_459_cast_fp16")];
+            tensor<int32, [3]> input_933_perm_0 = const()[name = string("input_933_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_935_pad_type_0 = const()[name = string("input_935_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_935_strides_0 = const()[name = string("input_935_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_935_pad_0 = const()[name = string("input_935_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_935_dilations_0 = const()[name = string("input_935_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_935_groups_0 = const()[name = string("input_935_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353329280))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355426496))))[name = string("encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_933_cast_fp16 = transpose(perm = input_933_perm_0, x = x_459_cast_fp16)[name = string("transpose_204")];
+            tensor<fp16, [1, 2048, 56]> input_935_cast_fp16 = conv(dilations = input_935_dilations_0, groups = input_935_groups_0, pad = input_935_pad_0, pad_type = input_935_pad_type_0, strides = input_935_strides_0, weight = encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_933_cast_fp16)[name = string("input_935_cast_fp16")];
+            int32 x_461_split_num_splits_0 = const()[name = string("x_461_split_num_splits_0"), val = int32(2)];
+            int32 x_461_split_axis_0 = const()[name = string("x_461_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_461_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_461_split_cast_fp16_1 = split(axis = x_461_split_axis_0, num_splits = x_461_split_num_splits_0, x = input_935_cast_fp16)[name = string("x_461_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_461_split_1_sigmoid_cast_fp16 = sigmoid(x = x_461_split_cast_fp16_1)[name = string("x_461_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_461_cast_fp16 = mul(x = x_461_split_cast_fp16_0, y = x_461_split_1_sigmoid_cast_fp16)[name = string("x_461_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_937_cast_fp16 = select(a = var_43_to_fp16, b = x_461_cast_fp16, cond = var_574)[name = string("input_937_cast_fp16")];
+            bool new_x_71_interleave_0 = const()[name = string("new_x_71_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_71_cast_fp16 = concat(axis = var_58, interleave = new_x_71_interleave_0, values = (cache_71_cast_fp16, input_937_cast_fp16))[name = string("new_x_71_cast_fp16")];
+            tensor<int32, [3]> var_4208_begin_0 = const()[name = string("op_4208_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_4208_end_0 = const()[name = string("op_4208_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_4208_end_mask_0 = const()[name = string("op_4208_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4208_cast_fp16 = slice_by_index(begin = var_4208_begin_0, end = var_4208_end_0, end_mask = var_4208_end_mask_0, x = new_x_71_cast_fp16)[name = string("op_4208_cast_fp16")];
+            string x_463_pad_type_0 = const()[name = string("x_463_pad_type_0"), val = string("valid")];
+            int32 x_463_groups_0 = const()[name = string("x_463_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_463_strides_0 = const()[name = string("x_463_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_463_pad_0 = const()[name = string("x_463_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_463_dilations_0 = const()[name = string("x_463_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355430656))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355439936))))[name = string("encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_463_cast_fp16 = conv(dilations = x_463_dilations_0, groups = x_463_groups_0, pad = x_463_pad_0, pad_type = x_463_pad_type_0, strides = x_463_strides_0, weight = encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_71_cast_fp16)[name = string("x_463_cast_fp16")];
+            tensor<int32, [3]> input_939_perm_0 = const()[name = string("input_939_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_465_axes_0 = const()[name = string("x_465_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355442048)))];
+            tensor<fp16, [1024]> encoder_layers_17_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355444160)))];
+            tensor<fp16, [1, 56, 1024]> input_939_cast_fp16 = transpose(perm = input_939_perm_0, x = x_463_cast_fp16)[name = string("transpose_203")];
+            tensor<fp16, [1, 56, 1024]> x_465_cast_fp16 = layer_norm(axes = x_465_axes_0, beta = encoder_layers_17_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_17_conv_batch_norm_weight_to_fp16, x = input_939_cast_fp16)[name = string("x_465_cast_fp16")];
+            tensor<int32, [3]> input_941_perm_0 = const()[name = string("input_941_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_941_cast_fp16 = transpose(perm = input_941_perm_0, x = x_465_cast_fp16)[name = string("transpose_202")];
+            tensor<fp16, [1, 1024, 56]> input_943_cast_fp16 = silu(x = input_941_cast_fp16)[name = string("input_943_cast_fp16")];
+            string x_467_pad_type_0 = const()[name = string("x_467_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_467_strides_0 = const()[name = string("x_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_467_pad_0 = const()[name = string("x_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_467_dilations_0 = const()[name = string("x_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_467_groups_0 = const()[name = string("x_467_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355446272))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356494912))))[name = string("encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_467_cast_fp16 = conv(dilations = x_467_dilations_0, groups = x_467_groups_0, pad = x_467_pad_0, pad_type = x_467_pad_type_0, strides = x_467_strides_0, weight = encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_943_cast_fp16)[name = string("x_467_cast_fp16")];
+            tensor<int32, [3]> input_945_perm_0 = const()[name = string("input_945_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_945_cast_fp16 = transpose(perm = input_945_perm_0, x = x_467_cast_fp16)[name = string("transpose_201")];
+            tensor<fp16, [1, 56, 1024]> input_947_cast_fp16 = add(x = input_931_cast_fp16, y = input_945_cast_fp16)[name = string("input_947_cast_fp16")];
+            tensor<int32, [1]> input_949_axes_0 = const()[name = string("input_949_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356497024)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356499136)))];
+            tensor<fp16, [1, 56, 1024]> input_949_cast_fp16 = layer_norm(axes = input_949_axes_0, beta = encoder_layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_17_norm_feed_forward2_weight_to_fp16, x = input_947_cast_fp16)[name = string("input_949_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356501248))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359647040))))[name = string("encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_17_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359647232)))];
+            tensor<fp16, [1, 56, 4096]> linear_161_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("linear_161_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_953_cast_fp16 = silu(x = linear_161_cast_fp16)[name = string("input_953_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359655488))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362801280))))[name = string("encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362801472)))];
+            tensor<fp16, [1, 56, 1024]> linear_162_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized, x = input_953_cast_fp16)[name = string("linear_162_cast_fp16")];
+            fp16 var_4251_to_fp16 = const()[name = string("op_4251_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4252_cast_fp16 = mul(x = linear_162_cast_fp16, y = var_4251_to_fp16)[name = string("op_4252_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_959_cast_fp16 = add(x = input_947_cast_fp16, y = var_4252_cast_fp16)[name = string("input_959_cast_fp16")];
+            tensor<int32, [1]> input_961_axes_0 = const()[name = string("input_961_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362803584)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362805696)))];
+            tensor<fp16, [1, 56, 1024]> input_961_cast_fp16 = layer_norm(axes = input_961_axes_0, beta = encoder_layers_17_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_17_norm_out_weight_to_fp16, x = input_959_cast_fp16)[name = string("input_961_cast_fp16")];
+            tensor<int32, [4]> cache_73_begin_0 = const()[name = string("cache_73_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> cache_73_end_0 = const()[name = string("cache_73_end_0"), val = tensor<int32, [4]>([19, 1, 42, 1024])];
+            tensor<bool, [4]> cache_73_end_mask_0 = const()[name = string("cache_73_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_73_squeeze_mask_0 = const()[name = string("cache_73_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_73_cast_fp16 = slice_by_index(begin = cache_73_begin_0, end = cache_73_end_0, end_mask = cache_73_end_mask_0, squeeze_mask = cache_73_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_73_cast_fp16")];
+            tensor<int32, [4]> cache_75_begin_0 = const()[name = string("cache_75_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> cache_75_end_0 = const()[name = string("cache_75_end_0"), val = tensor<int32, [4]>([19, 1, 1024, 8])];
+            tensor<bool, [4]> cache_75_end_mask_0 = const()[name = string("cache_75_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_75_squeeze_mask_0 = const()[name = string("cache_75_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_75_cast_fp16 = slice_by_index(begin = cache_75_begin_0, end = cache_75_end_0, end_mask = cache_75_end_mask_0, squeeze_mask = cache_75_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_75_cast_fp16")];
+            tensor<int32, [1]> input_963_axes_0 = const()[name = string("input_963_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362807808)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362809920)))];
+            tensor<fp16, [1, 56, 1024]> input_963_cast_fp16 = layer_norm(axes = input_963_axes_0, beta = encoder_layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_18_norm_feed_forward1_weight_to_fp16, x = input_961_cast_fp16)[name = string("input_963_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362812032))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365957824))))[name = string("encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_18_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365958016)))];
+            tensor<fp16, [1, 56, 4096]> linear_163_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized, x = input_963_cast_fp16)[name = string("linear_163_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_967_cast_fp16 = silu(x = linear_163_cast_fp16)[name = string("input_967_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(365966272))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369112064))))[name = string("encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369112256)))];
+            tensor<fp16, [1, 56, 1024]> linear_164_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized, x = input_967_cast_fp16)[name = string("linear_164_cast_fp16")];
+            fp16 var_4288_to_fp16 = const()[name = string("op_4288_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4289_cast_fp16 = mul(x = linear_164_cast_fp16, y = var_4288_to_fp16)[name = string("op_4289_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_973_cast_fp16 = add(x = input_961_cast_fp16, y = var_4289_cast_fp16)[name = string("input_973_cast_fp16")];
+            tensor<int32, [1]> key_37_axes_0 = const()[name = string("key_37_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369114368)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369116480)))];
+            tensor<fp16, [1, 56, 1024]> key_37_cast_fp16 = layer_norm(axes = key_37_axes_0, beta = encoder_layers_18_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_18_norm_self_att_weight_to_fp16, x = input_973_cast_fp16)[name = string("key_37_cast_fp16")];
+            bool input_975_interleave_0 = const()[name = string("input_975_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_975_cast_fp16 = concat(axis = var_67, interleave = input_975_interleave_0, values = (cache_73_cast_fp16, key_37_cast_fp16))[name = string("input_975_cast_fp16")];
+            bool var_4317_interleave_0 = const()[name = string("op_4317_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_4317_cast_fp16 = concat(axis = var_67, interleave = var_4317_interleave_0, values = key_37_cast_fp16)[name = string("op_4317_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369118592))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369905088))))[name = string("encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369905280)))];
+            tensor<fp16, [1, 56, 1024]> linear_165_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized, x = key_37_cast_fp16)[name = string("linear_165_cast_fp16")];
+            tensor<int32, [4]> var_4322 = const()[name = string("op_4322"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_109_cast_fp16 = reshape(shape = var_4322, x = linear_165_cast_fp16)[name = string("q_109_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369907392))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370693888))))[name = string("encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370694080)))];
+            tensor<fp16, [1, 98, 1024]> linear_166_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_166_cast_fp16")];
+            tensor<int32, [4]> var_4327 = const()[name = string("op_4327"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_73_cast_fp16 = reshape(shape = var_4327, x = linear_166_cast_fp16)[name = string("k_73_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370696192))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371482688))))[name = string("encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371482880)))];
+            tensor<fp16, [1, 98, 1024]> linear_167_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_167_cast_fp16")];
+            tensor<int32, [4]> var_4332 = const()[name = string("op_4332"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_37_cast_fp16 = reshape(shape = var_4332, x = linear_167_cast_fp16)[name = string("v_37_cast_fp16")];
+            tensor<int32, [4]> value_45_perm_0 = const()[name = string("value_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_18_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371484992)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4345_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_u_to_fp16)[name = string("op_4345_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_18_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371487104)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4347_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_v_to_fp16)[name = string("op_4347_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_37_perm_0 = const()[name = string("q_with_bias_v_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_475_transpose_x_0 = const()[name = string("x_475_transpose_x_0"), val = bool(false)];
+            bool x_475_transpose_y_0 = const()[name = string("x_475_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_4349_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371489216))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371688960))))[name = string("op_4349_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_37_cast_fp16 = transpose(perm = q_with_bias_v_37_perm_0, x = var_4347_cast_fp16)[name = string("transpose_200")];
+            tensor<fp16, [1, 8, 56, 195]> x_475_cast_fp16 = matmul(transpose_x = x_475_transpose_x_0, transpose_y = x_475_transpose_y_0, x = q_with_bias_v_37_cast_fp16, y = op_4349_to_fp16_quantized)[name = string("x_475_cast_fp16")];
+            tensor<int32, [8]> x_477_pad_0 = const()[name = string("x_477_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_477_mode_0 = const()[name = string("x_477_mode_0"), val = string("constant")];
+            fp16 const_313_to_fp16 = const()[name = string("const_313_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_477_cast_fp16 = pad(constant_val = const_313_to_fp16, mode = x_477_mode_0, pad = x_477_pad_0, x = x_475_cast_fp16)[name = string("x_477_cast_fp16")];
+            tensor<int32, [4]> var_4357 = const()[name = string("op_4357"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_479_cast_fp16 = reshape(shape = var_4357, x = x_477_cast_fp16)[name = string("x_479_cast_fp16")];
+            tensor<int32, [4]> var_4361_begin_0 = const()[name = string("op_4361_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4361_end_0 = const()[name = string("op_4361_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_4361_end_mask_0 = const()[name = string("op_4361_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_4361_cast_fp16 = slice_by_index(begin = var_4361_begin_0, end = var_4361_end_0, end_mask = var_4361_end_mask_0, x = x_479_cast_fp16)[name = string("op_4361_cast_fp16")];
+            tensor<int32, [4]> var_4362 = const()[name = string("op_4362"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_73_cast_fp16 = reshape(shape = var_4362, x = var_4361_cast_fp16)[name = string("matrix_bd_73_cast_fp16")];
+            bool matrix_ac_37_transpose_x_0 = const()[name = string("matrix_ac_37_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_37_transpose_y_0 = const()[name = string("matrix_ac_37_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_132_perm_0 = const()[name = string("transpose_132_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_133_perm_0 = const()[name = string("transpose_133_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_133 = transpose(perm = transpose_133_perm_0, x = k_73_cast_fp16)[name = string("transpose_198")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_132 = transpose(perm = transpose_132_perm_0, x = var_4345_cast_fp16)[name = string("transpose_199")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_37_cast_fp16 = matmul(transpose_x = matrix_ac_37_transpose_x_0, transpose_y = matrix_ac_37_transpose_y_0, x = transpose_132, y = transpose_133)[name = string("matrix_ac_37_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_75_begin_0 = const()[name = string("matrix_bd_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_75_end_0 = const()[name = string("matrix_bd_75_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_75_end_mask_0 = const()[name = string("matrix_bd_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_75_cast_fp16 = slice_by_index(begin = matrix_bd_75_begin_0, end = matrix_bd_75_end_0, end_mask = matrix_bd_75_end_mask_0, x = matrix_bd_73_cast_fp16)[name = string("matrix_bd_75_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_4371_cast_fp16 = add(x = matrix_ac_37_cast_fp16, y = matrix_bd_75_cast_fp16)[name = string("op_4371_cast_fp16")];
+            fp16 _inversed_scores_73_y_0_to_fp16 = const()[name = string("_inversed_scores_73_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_73_cast_fp16 = mul(x = var_4371_cast_fp16, y = _inversed_scores_73_y_0_to_fp16)[name = string("_inversed_scores_73_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_75_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_73_cast_fp16, cond = mask_11)[name = string("scores_75_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_4377_cast_fp16 = softmax(axis = var_58, x = scores_75_cast_fp16)[name = string("op_4377_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_977_cast_fp16 = select(a = var_43_to_fp16, b = var_4377_cast_fp16, cond = mask_11)[name = string("input_977_cast_fp16")];
+            bool x_481_transpose_x_0 = const()[name = string("x_481_transpose_x_0"), val = bool(false)];
+            bool x_481_transpose_y_0 = const()[name = string("x_481_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = v_37_cast_fp16)[name = string("transpose_197")];
+            tensor<fp16, [1, 8, 56, 128]> x_481_cast_fp16 = matmul(transpose_x = x_481_transpose_x_0, transpose_y = x_481_transpose_y_0, x = input_977_cast_fp16, y = value_45_cast_fp16)[name = string("x_481_cast_fp16")];
+            tensor<int32, [4]> var_4381_perm_0 = const()[name = string("op_4381_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4382 = const()[name = string("op_4382"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_4381_cast_fp16 = transpose(perm = var_4381_perm_0, x = x_481_cast_fp16)[name = string("transpose_196")];
+            tensor<fp16, [1, 56, 1024]> input_979_cast_fp16 = reshape(shape = var_4382, x = var_4381_cast_fp16)[name = string("input_979_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371689472))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372738112))))[name = string("encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372740224)))];
+            tensor<fp16, [1, 56, 1024]> linear_169_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized, x = input_979_cast_fp16)[name = string("linear_169_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_983_cast_fp16 = add(x = input_973_cast_fp16, y = linear_169_cast_fp16)[name = string("input_983_cast_fp16")];
+            tensor<int32, [1]> x_485_axes_0 = const()[name = string("x_485_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372742336)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372744448)))];
+            tensor<fp16, [1, 56, 1024]> x_485_cast_fp16 = layer_norm(axes = x_485_axes_0, beta = encoder_layers_18_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_18_norm_conv_weight_to_fp16, x = input_983_cast_fp16)[name = string("x_485_cast_fp16")];
+            tensor<int32, [3]> input_985_perm_0 = const()[name = string("input_985_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_987_pad_type_0 = const()[name = string("input_987_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_987_strides_0 = const()[name = string("input_987_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_987_pad_0 = const()[name = string("input_987_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_987_dilations_0 = const()[name = string("input_987_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_987_groups_0 = const()[name = string("input_987_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372746560))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374843776))))[name = string("encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_985_cast_fp16 = transpose(perm = input_985_perm_0, x = x_485_cast_fp16)[name = string("transpose_195")];
+            tensor<fp16, [1, 2048, 56]> input_987_cast_fp16 = conv(dilations = input_987_dilations_0, groups = input_987_groups_0, pad = input_987_pad_0, pad_type = input_987_pad_type_0, strides = input_987_strides_0, weight = encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_985_cast_fp16)[name = string("input_987_cast_fp16")];
+            int32 x_487_split_num_splits_0 = const()[name = string("x_487_split_num_splits_0"), val = int32(2)];
+            int32 x_487_split_axis_0 = const()[name = string("x_487_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_487_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_487_split_cast_fp16_1 = split(axis = x_487_split_axis_0, num_splits = x_487_split_num_splits_0, x = input_987_cast_fp16)[name = string("x_487_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_487_split_1_sigmoid_cast_fp16 = sigmoid(x = x_487_split_cast_fp16_1)[name = string("x_487_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_487_cast_fp16 = mul(x = x_487_split_cast_fp16_0, y = x_487_split_1_sigmoid_cast_fp16)[name = string("x_487_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_989_cast_fp16 = select(a = var_43_to_fp16, b = x_487_cast_fp16, cond = var_574)[name = string("input_989_cast_fp16")];
+            bool new_x_75_interleave_0 = const()[name = string("new_x_75_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_75_cast_fp16 = concat(axis = var_58, interleave = new_x_75_interleave_0, values = (cache_75_cast_fp16, input_989_cast_fp16))[name = string("new_x_75_cast_fp16")];
+            tensor<int32, [3]> var_4421_begin_0 = const()[name = string("op_4421_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_4421_end_0 = const()[name = string("op_4421_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_4421_end_mask_0 = const()[name = string("op_4421_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4421_cast_fp16 = slice_by_index(begin = var_4421_begin_0, end = var_4421_end_0, end_mask = var_4421_end_mask_0, x = new_x_75_cast_fp16)[name = string("op_4421_cast_fp16")];
+            string x_489_pad_type_0 = const()[name = string("x_489_pad_type_0"), val = string("valid")];
+            int32 x_489_groups_0 = const()[name = string("x_489_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_489_strides_0 = const()[name = string("x_489_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_489_pad_0 = const()[name = string("x_489_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_489_dilations_0 = const()[name = string("x_489_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374847936))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374857216))))[name = string("encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_489_cast_fp16 = conv(dilations = x_489_dilations_0, groups = x_489_groups_0, pad = x_489_pad_0, pad_type = x_489_pad_type_0, strides = x_489_strides_0, weight = encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_75_cast_fp16)[name = string("x_489_cast_fp16")];
+            tensor<int32, [3]> input_991_perm_0 = const()[name = string("input_991_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_491_axes_0 = const()[name = string("x_491_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374859328)))];
+            tensor<fp16, [1024]> encoder_layers_18_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374861440)))];
+            tensor<fp16, [1, 56, 1024]> input_991_cast_fp16 = transpose(perm = input_991_perm_0, x = x_489_cast_fp16)[name = string("transpose_194")];
+            tensor<fp16, [1, 56, 1024]> x_491_cast_fp16 = layer_norm(axes = x_491_axes_0, beta = encoder_layers_18_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_18_conv_batch_norm_weight_to_fp16, x = input_991_cast_fp16)[name = string("x_491_cast_fp16")];
+            tensor<int32, [3]> input_993_perm_0 = const()[name = string("input_993_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_993_cast_fp16 = transpose(perm = input_993_perm_0, x = x_491_cast_fp16)[name = string("transpose_193")];
+            tensor<fp16, [1, 1024, 56]> input_995_cast_fp16 = silu(x = input_993_cast_fp16)[name = string("input_995_cast_fp16")];
+            string x_493_pad_type_0 = const()[name = string("x_493_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_493_strides_0 = const()[name = string("x_493_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_493_pad_0 = const()[name = string("x_493_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_493_dilations_0 = const()[name = string("x_493_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_493_groups_0 = const()[name = string("x_493_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374863552))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375912192))))[name = string("encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_493_cast_fp16 = conv(dilations = x_493_dilations_0, groups = x_493_groups_0, pad = x_493_pad_0, pad_type = x_493_pad_type_0, strides = x_493_strides_0, weight = encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_995_cast_fp16)[name = string("x_493_cast_fp16")];
+            tensor<int32, [3]> input_997_perm_0 = const()[name = string("input_997_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_997_cast_fp16 = transpose(perm = input_997_perm_0, x = x_493_cast_fp16)[name = string("transpose_192")];
+            tensor<fp16, [1, 56, 1024]> input_999_cast_fp16 = add(x = input_983_cast_fp16, y = input_997_cast_fp16)[name = string("input_999_cast_fp16")];
+            tensor<int32, [1]> input_1001_axes_0 = const()[name = string("input_1001_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375914304)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375916416)))];
+            tensor<fp16, [1, 56, 1024]> input_1001_cast_fp16 = layer_norm(axes = input_1001_axes_0, beta = encoder_layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_18_norm_feed_forward2_weight_to_fp16, x = input_999_cast_fp16)[name = string("input_1001_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(375918528))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380112896))))[name = string("encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_18_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380121152)))];
+            tensor<fp16, [1, 56, 4096]> linear_170_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1001_cast_fp16)[name = string("linear_170_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1005_cast_fp16 = silu(x = linear_170_cast_fp16)[name = string("input_1005_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380129408))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384323776))))[name = string("encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_18_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384325888)))];
+            tensor<fp16, [1, 56, 1024]> linear_171_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1005_cast_fp16)[name = string("linear_171_cast_fp16")];
+            fp16 var_4464_to_fp16 = const()[name = string("op_4464_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4465_cast_fp16 = mul(x = linear_171_cast_fp16, y = var_4464_to_fp16)[name = string("op_4465_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1011_cast_fp16 = add(x = input_999_cast_fp16, y = var_4465_cast_fp16)[name = string("input_1011_cast_fp16")];
+            tensor<int32, [1]> input_1013_axes_0 = const()[name = string("input_1013_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384328000)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384330112)))];
+            tensor<fp16, [1, 56, 1024]> input_1013_cast_fp16 = layer_norm(axes = input_1013_axes_0, beta = encoder_layers_18_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_18_norm_out_weight_to_fp16, x = input_1011_cast_fp16)[name = string("input_1013_cast_fp16")];
+            tensor<int32, [4]> cache_77_begin_0 = const()[name = string("cache_77_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
+            tensor<int32, [4]> cache_77_end_0 = const()[name = string("cache_77_end_0"), val = tensor<int32, [4]>([20, 1, 42, 1024])];
+            tensor<bool, [4]> cache_77_end_mask_0 = const()[name = string("cache_77_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_77_squeeze_mask_0 = const()[name = string("cache_77_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_77_cast_fp16 = slice_by_index(begin = cache_77_begin_0, end = cache_77_end_0, end_mask = cache_77_end_mask_0, squeeze_mask = cache_77_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_77_cast_fp16")];
+            tensor<int32, [4]> cache_79_begin_0 = const()[name = string("cache_79_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
+            tensor<int32, [4]> cache_79_end_0 = const()[name = string("cache_79_end_0"), val = tensor<int32, [4]>([20, 1, 1024, 8])];
+            tensor<bool, [4]> cache_79_end_mask_0 = const()[name = string("cache_79_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_79_squeeze_mask_0 = const()[name = string("cache_79_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_79_cast_fp16 = slice_by_index(begin = cache_79_begin_0, end = cache_79_end_0, end_mask = cache_79_end_mask_0, squeeze_mask = cache_79_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_79_cast_fp16")];
+            tensor<int32, [1]> input_1015_axes_0 = const()[name = string("input_1015_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384332224)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384334336)))];
+            tensor<fp16, [1, 56, 1024]> input_1015_cast_fp16 = layer_norm(axes = input_1015_axes_0, beta = encoder_layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_19_norm_feed_forward1_weight_to_fp16, x = input_1013_cast_fp16)[name = string("input_1015_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384336448))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388530816))))[name = string("encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_19_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388539072)))];
+            tensor<fp16, [1, 56, 4096]> linear_172_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1015_cast_fp16)[name = string("linear_172_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1019_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("input_1019_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(388547328))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392741696))))[name = string("encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392743808)))];
+            tensor<fp16, [1, 56, 1024]> linear_173_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1019_cast_fp16)[name = string("linear_173_cast_fp16")];
+            fp16 var_4501_to_fp16 = const()[name = string("op_4501_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4502_cast_fp16 = mul(x = linear_173_cast_fp16, y = var_4501_to_fp16)[name = string("op_4502_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1025_cast_fp16 = add(x = input_1013_cast_fp16, y = var_4502_cast_fp16)[name = string("input_1025_cast_fp16")];
+            tensor<int32, [1]> key_39_axes_0 = const()[name = string("key_39_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392745920)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392748032)))];
+            tensor<fp16, [1, 56, 1024]> key_39_cast_fp16 = layer_norm(axes = key_39_axes_0, beta = encoder_layers_19_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_19_norm_self_att_weight_to_fp16, x = input_1025_cast_fp16)[name = string("key_39_cast_fp16")];
+            bool input_1027_interleave_0 = const()[name = string("input_1027_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_1027_cast_fp16 = concat(axis = var_67, interleave = input_1027_interleave_0, values = (cache_77_cast_fp16, key_39_cast_fp16))[name = string("input_1027_cast_fp16")];
+            bool var_4530_interleave_0 = const()[name = string("op_4530_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_4530_cast_fp16 = concat(axis = var_67, interleave = var_4530_interleave_0, values = key_39_cast_fp16)[name = string("op_4530_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392750144))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393798784))))[name = string("encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393800896)))];
+            tensor<fp16, [1, 56, 1024]> linear_174_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized, x = key_39_cast_fp16)[name = string("linear_174_cast_fp16")];
+            tensor<int32, [4]> var_4535 = const()[name = string("op_4535"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_115_cast_fp16 = reshape(shape = var_4535, x = linear_174_cast_fp16)[name = string("q_115_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393803008))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394851648))))[name = string("encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394853760)))];
+            tensor<fp16, [1, 98, 1024]> linear_175_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_175_cast_fp16")];
+            tensor<int32, [4]> var_4540 = const()[name = string("op_4540"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_77_cast_fp16 = reshape(shape = var_4540, x = linear_175_cast_fp16)[name = string("k_77_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394855872))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395904512))))[name = string("encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395906624)))];
+            tensor<fp16, [1, 98, 1024]> linear_176_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_176_cast_fp16")];
+            tensor<int32, [4]> var_4545 = const()[name = string("op_4545"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_39_cast_fp16 = reshape(shape = var_4545, x = linear_176_cast_fp16)[name = string("v_39_cast_fp16")];
+            tensor<int32, [4]> value_47_perm_0 = const()[name = string("value_47_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_19_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395908736)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4558_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_u_to_fp16)[name = string("op_4558_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_19_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395910848)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4560_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_v_to_fp16)[name = string("op_4560_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_39_perm_0 = const()[name = string("q_with_bias_v_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_501_transpose_x_0 = const()[name = string("x_501_transpose_x_0"), val = bool(false)];
+            bool x_501_transpose_y_0 = const()[name = string("x_501_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_4562_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395912960))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396112704))))[name = string("op_4562_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_39_cast_fp16 = transpose(perm = q_with_bias_v_39_perm_0, x = var_4560_cast_fp16)[name = string("transpose_191")];
+            tensor<fp16, [1, 8, 56, 195]> x_501_cast_fp16 = matmul(transpose_x = x_501_transpose_x_0, transpose_y = x_501_transpose_y_0, x = q_with_bias_v_39_cast_fp16, y = op_4562_to_fp16_quantized)[name = string("x_501_cast_fp16")];
+            tensor<int32, [8]> x_503_pad_0 = const()[name = string("x_503_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_503_mode_0 = const()[name = string("x_503_mode_0"), val = string("constant")];
+            fp16 const_326_to_fp16 = const()[name = string("const_326_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_503_cast_fp16 = pad(constant_val = const_326_to_fp16, mode = x_503_mode_0, pad = x_503_pad_0, x = x_501_cast_fp16)[name = string("x_503_cast_fp16")];
+            tensor<int32, [4]> var_4570 = const()[name = string("op_4570"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_505_cast_fp16 = reshape(shape = var_4570, x = x_503_cast_fp16)[name = string("x_505_cast_fp16")];
+            tensor<int32, [4]> var_4574_begin_0 = const()[name = string("op_4574_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4574_end_0 = const()[name = string("op_4574_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_4574_end_mask_0 = const()[name = string("op_4574_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_4574_cast_fp16 = slice_by_index(begin = var_4574_begin_0, end = var_4574_end_0, end_mask = var_4574_end_mask_0, x = x_505_cast_fp16)[name = string("op_4574_cast_fp16")];
+            tensor<int32, [4]> var_4575 = const()[name = string("op_4575"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_77_cast_fp16 = reshape(shape = var_4575, x = var_4574_cast_fp16)[name = string("matrix_bd_77_cast_fp16")];
+            bool matrix_ac_39_transpose_x_0 = const()[name = string("matrix_ac_39_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_39_transpose_y_0 = const()[name = string("matrix_ac_39_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_134_perm_0 = const()[name = string("transpose_134_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_135_perm_0 = const()[name = string("transpose_135_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_135 = transpose(perm = transpose_135_perm_0, x = k_77_cast_fp16)[name = string("transpose_189")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_134 = transpose(perm = transpose_134_perm_0, x = var_4558_cast_fp16)[name = string("transpose_190")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_39_cast_fp16 = matmul(transpose_x = matrix_ac_39_transpose_x_0, transpose_y = matrix_ac_39_transpose_y_0, x = transpose_134, y = transpose_135)[name = string("matrix_ac_39_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_79_begin_0 = const()[name = string("matrix_bd_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_79_end_0 = const()[name = string("matrix_bd_79_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_79_end_mask_0 = const()[name = string("matrix_bd_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_79_cast_fp16 = slice_by_index(begin = matrix_bd_79_begin_0, end = matrix_bd_79_end_0, end_mask = matrix_bd_79_end_mask_0, x = matrix_bd_77_cast_fp16)[name = string("matrix_bd_79_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_4584_cast_fp16 = add(x = matrix_ac_39_cast_fp16, y = matrix_bd_79_cast_fp16)[name = string("op_4584_cast_fp16")];
+            fp16 _inversed_scores_77_y_0_to_fp16 = const()[name = string("_inversed_scores_77_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_77_cast_fp16 = mul(x = var_4584_cast_fp16, y = _inversed_scores_77_y_0_to_fp16)[name = string("_inversed_scores_77_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_79_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_77_cast_fp16, cond = mask_11)[name = string("scores_79_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_4590_cast_fp16 = softmax(axis = var_58, x = scores_79_cast_fp16)[name = string("op_4590_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_1029_cast_fp16 = select(a = var_43_to_fp16, b = var_4590_cast_fp16, cond = mask_11)[name = string("input_1029_cast_fp16")];
+            bool x_507_transpose_x_0 = const()[name = string("x_507_transpose_x_0"), val = bool(false)];
+            bool x_507_transpose_y_0 = const()[name = string("x_507_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = v_39_cast_fp16)[name = string("transpose_188")];
+            tensor<fp16, [1, 8, 56, 128]> x_507_cast_fp16 = matmul(transpose_x = x_507_transpose_x_0, transpose_y = x_507_transpose_y_0, x = input_1029_cast_fp16, y = value_47_cast_fp16)[name = string("x_507_cast_fp16")];
+            tensor<int32, [4]> var_4594_perm_0 = const()[name = string("op_4594_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4595 = const()[name = string("op_4595"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_4594_cast_fp16 = transpose(perm = var_4594_perm_0, x = x_507_cast_fp16)[name = string("transpose_187")];
+            tensor<fp16, [1, 56, 1024]> input_1031_cast_fp16 = reshape(shape = var_4595, x = var_4594_cast_fp16)[name = string("input_1031_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396113216))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397161856))))[name = string("encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397163968)))];
+            tensor<fp16, [1, 56, 1024]> linear_178_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized, x = input_1031_cast_fp16)[name = string("linear_178_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1035_cast_fp16 = add(x = input_1025_cast_fp16, y = linear_178_cast_fp16)[name = string("input_1035_cast_fp16")];
+            tensor<int32, [1]> x_511_axes_0 = const()[name = string("x_511_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397166080)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397168192)))];
+            tensor<fp16, [1, 56, 1024]> x_511_cast_fp16 = layer_norm(axes = x_511_axes_0, beta = encoder_layers_19_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_19_norm_conv_weight_to_fp16, x = input_1035_cast_fp16)[name = string("x_511_cast_fp16")];
+            tensor<int32, [3]> input_1037_perm_0 = const()[name = string("input_1037_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1039_pad_type_0 = const()[name = string("input_1039_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1039_strides_0 = const()[name = string("input_1039_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1039_pad_0 = const()[name = string("input_1039_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1039_dilations_0 = const()[name = string("input_1039_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1039_groups_0 = const()[name = string("input_1039_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397170304))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399267520))))[name = string("encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_1037_cast_fp16 = transpose(perm = input_1037_perm_0, x = x_511_cast_fp16)[name = string("transpose_186")];
+            tensor<fp16, [1, 2048, 56]> input_1039_cast_fp16 = conv(dilations = input_1039_dilations_0, groups = input_1039_groups_0, pad = input_1039_pad_0, pad_type = input_1039_pad_type_0, strides = input_1039_strides_0, weight = encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1037_cast_fp16)[name = string("input_1039_cast_fp16")];
+            int32 x_513_split_num_splits_0 = const()[name = string("x_513_split_num_splits_0"), val = int32(2)];
+            int32 x_513_split_axis_0 = const()[name = string("x_513_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_513_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_513_split_cast_fp16_1 = split(axis = x_513_split_axis_0, num_splits = x_513_split_num_splits_0, x = input_1039_cast_fp16)[name = string("x_513_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_513_split_1_sigmoid_cast_fp16 = sigmoid(x = x_513_split_cast_fp16_1)[name = string("x_513_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_513_cast_fp16 = mul(x = x_513_split_cast_fp16_0, y = x_513_split_1_sigmoid_cast_fp16)[name = string("x_513_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_1041_cast_fp16 = select(a = var_43_to_fp16, b = x_513_cast_fp16, cond = var_574)[name = string("input_1041_cast_fp16")];
+            bool new_x_79_interleave_0 = const()[name = string("new_x_79_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_79_cast_fp16 = concat(axis = var_58, interleave = new_x_79_interleave_0, values = (cache_79_cast_fp16, input_1041_cast_fp16))[name = string("new_x_79_cast_fp16")];
+            tensor<int32, [3]> var_4634_begin_0 = const()[name = string("op_4634_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_4634_end_0 = const()[name = string("op_4634_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_4634_end_mask_0 = const()[name = string("op_4634_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4634_cast_fp16 = slice_by_index(begin = var_4634_begin_0, end = var_4634_end_0, end_mask = var_4634_end_mask_0, x = new_x_79_cast_fp16)[name = string("op_4634_cast_fp16")];
+            string x_515_pad_type_0 = const()[name = string("x_515_pad_type_0"), val = string("valid")];
+            int32 x_515_groups_0 = const()[name = string("x_515_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_515_strides_0 = const()[name = string("x_515_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_515_pad_0 = const()[name = string("x_515_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_515_dilations_0 = const()[name = string("x_515_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399271680))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399280960))))[name = string("encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_515_cast_fp16 = conv(dilations = x_515_dilations_0, groups = x_515_groups_0, pad = x_515_pad_0, pad_type = x_515_pad_type_0, strides = x_515_strides_0, weight = encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_79_cast_fp16)[name = string("x_515_cast_fp16")];
+            tensor<int32, [3]> input_1043_perm_0 = const()[name = string("input_1043_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_517_axes_0 = const()[name = string("x_517_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399283072)))];
+            tensor<fp16, [1024]> encoder_layers_19_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399285184)))];
+            tensor<fp16, [1, 56, 1024]> input_1043_cast_fp16 = transpose(perm = input_1043_perm_0, x = x_515_cast_fp16)[name = string("transpose_185")];
+            tensor<fp16, [1, 56, 1024]> x_517_cast_fp16 = layer_norm(axes = x_517_axes_0, beta = encoder_layers_19_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_19_conv_batch_norm_weight_to_fp16, x = input_1043_cast_fp16)[name = string("x_517_cast_fp16")];
+            tensor<int32, [3]> input_1045_perm_0 = const()[name = string("input_1045_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_1045_cast_fp16 = transpose(perm = input_1045_perm_0, x = x_517_cast_fp16)[name = string("transpose_184")];
+            tensor<fp16, [1, 1024, 56]> input_1047_cast_fp16 = silu(x = input_1045_cast_fp16)[name = string("input_1047_cast_fp16")];
+            string x_519_pad_type_0 = const()[name = string("x_519_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_519_strides_0 = const()[name = string("x_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_519_pad_0 = const()[name = string("x_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_519_dilations_0 = const()[name = string("x_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_519_groups_0 = const()[name = string("x_519_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399287296))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400335936))))[name = string("encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_519_cast_fp16 = conv(dilations = x_519_dilations_0, groups = x_519_groups_0, pad = x_519_pad_0, pad_type = x_519_pad_type_0, strides = x_519_strides_0, weight = encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1047_cast_fp16)[name = string("x_519_cast_fp16")];
+            tensor<int32, [3]> input_1049_perm_0 = const()[name = string("input_1049_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_1049_cast_fp16 = transpose(perm = input_1049_perm_0, x = x_519_cast_fp16)[name = string("transpose_183")];
+            tensor<fp16, [1, 56, 1024]> input_1051_cast_fp16 = add(x = input_1035_cast_fp16, y = input_1049_cast_fp16)[name = string("input_1051_cast_fp16")];
+            tensor<int32, [1]> input_1053_axes_0 = const()[name = string("input_1053_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400338048)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400340160)))];
+            tensor<fp16, [1, 56, 1024]> input_1053_cast_fp16 = layer_norm(axes = input_1053_axes_0, beta = encoder_layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_19_norm_feed_forward2_weight_to_fp16, x = input_1051_cast_fp16)[name = string("input_1053_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400342272))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404536640))))[name = string("encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_19_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404544896)))];
+            tensor<fp16, [1, 56, 4096]> linear_179_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1053_cast_fp16)[name = string("linear_179_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1057_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("input_1057_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(404553152))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408747520))))[name = string("encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408749632)))];
+            tensor<fp16, [1, 56, 1024]> linear_180_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1057_cast_fp16)[name = string("linear_180_cast_fp16")];
+            fp16 var_4677_to_fp16 = const()[name = string("op_4677_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4678_cast_fp16 = mul(x = linear_180_cast_fp16, y = var_4677_to_fp16)[name = string("op_4678_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1063_cast_fp16 = add(x = input_1051_cast_fp16, y = var_4678_cast_fp16)[name = string("input_1063_cast_fp16")];
+            tensor<int32, [1]> input_1065_axes_0 = const()[name = string("input_1065_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408751744)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408753856)))];
+            tensor<fp16, [1, 56, 1024]> input_1065_cast_fp16 = layer_norm(axes = input_1065_axes_0, beta = encoder_layers_19_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_19_norm_out_weight_to_fp16, x = input_1063_cast_fp16)[name = string("input_1065_cast_fp16")];
+            tensor<int32, [4]> cache_81_begin_0 = const()[name = string("cache_81_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
+            tensor<int32, [4]> cache_81_end_0 = const()[name = string("cache_81_end_0"), val = tensor<int32, [4]>([21, 1, 42, 1024])];
+            tensor<bool, [4]> cache_81_end_mask_0 = const()[name = string("cache_81_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_81_squeeze_mask_0 = const()[name = string("cache_81_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_81_cast_fp16 = slice_by_index(begin = cache_81_begin_0, end = cache_81_end_0, end_mask = cache_81_end_mask_0, squeeze_mask = cache_81_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_81_cast_fp16")];
+            tensor<int32, [4]> cache_83_begin_0 = const()[name = string("cache_83_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
+            tensor<int32, [4]> cache_83_end_0 = const()[name = string("cache_83_end_0"), val = tensor<int32, [4]>([21, 1, 1024, 8])];
+            tensor<bool, [4]> cache_83_end_mask_0 = const()[name = string("cache_83_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_83_squeeze_mask_0 = const()[name = string("cache_83_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_83_cast_fp16 = slice_by_index(begin = cache_83_begin_0, end = cache_83_end_0, end_mask = cache_83_end_mask_0, squeeze_mask = cache_83_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_83_cast_fp16")];
+            tensor<int32, [1]> input_1067_axes_0 = const()[name = string("input_1067_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408755968)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408758080)))];
+            tensor<fp16, [1, 56, 1024]> input_1067_cast_fp16 = layer_norm(axes = input_1067_axes_0, beta = encoder_layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_20_norm_feed_forward1_weight_to_fp16, x = input_1065_cast_fp16)[name = string("input_1067_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408760192))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412954560))))[name = string("encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_20_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412962816)))];
+            tensor<fp16, [1, 56, 4096]> linear_181_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1067_cast_fp16)[name = string("linear_181_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1071_cast_fp16 = silu(x = linear_181_cast_fp16)[name = string("input_1071_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412971072))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417165440))))[name = string("encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417167552)))];
+            tensor<fp16, [1, 56, 1024]> linear_182_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1071_cast_fp16)[name = string("linear_182_cast_fp16")];
+            fp16 var_4714_to_fp16 = const()[name = string("op_4714_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4715_cast_fp16 = mul(x = linear_182_cast_fp16, y = var_4714_to_fp16)[name = string("op_4715_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1077_cast_fp16 = add(x = input_1065_cast_fp16, y = var_4715_cast_fp16)[name = string("input_1077_cast_fp16")];
+            tensor<int32, [1]> key_41_axes_0 = const()[name = string("key_41_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417169664)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417171776)))];
+            tensor<fp16, [1, 56, 1024]> key_41_cast_fp16 = layer_norm(axes = key_41_axes_0, beta = encoder_layers_20_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_20_norm_self_att_weight_to_fp16, x = input_1077_cast_fp16)[name = string("key_41_cast_fp16")];
+            bool input_1079_interleave_0 = const()[name = string("input_1079_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_1079_cast_fp16 = concat(axis = var_67, interleave = input_1079_interleave_0, values = (cache_81_cast_fp16, key_41_cast_fp16))[name = string("input_1079_cast_fp16")];
+            bool var_4743_interleave_0 = const()[name = string("op_4743_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_4743_cast_fp16 = concat(axis = var_67, interleave = var_4743_interleave_0, values = key_41_cast_fp16)[name = string("op_4743_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417173888))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418222528))))[name = string("encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418224640)))];
+            tensor<fp16, [1, 56, 1024]> linear_183_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized, x = key_41_cast_fp16)[name = string("linear_183_cast_fp16")];
+            tensor<int32, [4]> var_4748 = const()[name = string("op_4748"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_121_cast_fp16 = reshape(shape = var_4748, x = linear_183_cast_fp16)[name = string("q_121_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418226752))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419275392))))[name = string("encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419277504)))];
+            tensor<fp16, [1, 98, 1024]> linear_184_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_184_cast_fp16")];
+            tensor<int32, [4]> var_4753 = const()[name = string("op_4753"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_81_cast_fp16 = reshape(shape = var_4753, x = linear_184_cast_fp16)[name = string("k_81_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419279616))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420328256))))[name = string("encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420330368)))];
+            tensor<fp16, [1, 98, 1024]> linear_185_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_185_cast_fp16")];
+            tensor<int32, [4]> var_4758 = const()[name = string("op_4758"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_41_cast_fp16 = reshape(shape = var_4758, x = linear_185_cast_fp16)[name = string("v_41_cast_fp16")];
+            tensor<int32, [4]> value_49_perm_0 = const()[name = string("value_49_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_20_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420332480)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4771_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_u_to_fp16)[name = string("op_4771_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_20_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420334592)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4773_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_v_to_fp16)[name = string("op_4773_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_41_perm_0 = const()[name = string("q_with_bias_v_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_527_transpose_x_0 = const()[name = string("x_527_transpose_x_0"), val = bool(false)];
+            bool x_527_transpose_y_0 = const()[name = string("x_527_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_4775_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420336704))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420536448))))[name = string("op_4775_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_41_cast_fp16 = transpose(perm = q_with_bias_v_41_perm_0, x = var_4773_cast_fp16)[name = string("transpose_182")];
+            tensor<fp16, [1, 8, 56, 195]> x_527_cast_fp16 = matmul(transpose_x = x_527_transpose_x_0, transpose_y = x_527_transpose_y_0, x = q_with_bias_v_41_cast_fp16, y = op_4775_to_fp16_quantized)[name = string("x_527_cast_fp16")];
+            tensor<int32, [8]> x_529_pad_0 = const()[name = string("x_529_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_529_mode_0 = const()[name = string("x_529_mode_0"), val = string("constant")];
+            fp16 const_339_to_fp16 = const()[name = string("const_339_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_529_cast_fp16 = pad(constant_val = const_339_to_fp16, mode = x_529_mode_0, pad = x_529_pad_0, x = x_527_cast_fp16)[name = string("x_529_cast_fp16")];
+            tensor<int32, [4]> var_4783 = const()[name = string("op_4783"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_531_cast_fp16 = reshape(shape = var_4783, x = x_529_cast_fp16)[name = string("x_531_cast_fp16")];
+            tensor<int32, [4]> var_4787_begin_0 = const()[name = string("op_4787_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4787_end_0 = const()[name = string("op_4787_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_4787_end_mask_0 = const()[name = string("op_4787_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_4787_cast_fp16 = slice_by_index(begin = var_4787_begin_0, end = var_4787_end_0, end_mask = var_4787_end_mask_0, x = x_531_cast_fp16)[name = string("op_4787_cast_fp16")];
+            tensor<int32, [4]> var_4788 = const()[name = string("op_4788"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_81_cast_fp16 = reshape(shape = var_4788, x = var_4787_cast_fp16)[name = string("matrix_bd_81_cast_fp16")];
+            bool matrix_ac_41_transpose_x_0 = const()[name = string("matrix_ac_41_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_41_transpose_y_0 = const()[name = string("matrix_ac_41_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_136_perm_0 = const()[name = string("transpose_136_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_137_perm_0 = const()[name = string("transpose_137_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_137 = transpose(perm = transpose_137_perm_0, x = k_81_cast_fp16)[name = string("transpose_180")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_136 = transpose(perm = transpose_136_perm_0, x = var_4771_cast_fp16)[name = string("transpose_181")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_41_cast_fp16 = matmul(transpose_x = matrix_ac_41_transpose_x_0, transpose_y = matrix_ac_41_transpose_y_0, x = transpose_136, y = transpose_137)[name = string("matrix_ac_41_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_83_begin_0 = const()[name = string("matrix_bd_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_83_end_0 = const()[name = string("matrix_bd_83_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_83_end_mask_0 = const()[name = string("matrix_bd_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_83_cast_fp16 = slice_by_index(begin = matrix_bd_83_begin_0, end = matrix_bd_83_end_0, end_mask = matrix_bd_83_end_mask_0, x = matrix_bd_81_cast_fp16)[name = string("matrix_bd_83_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_4797_cast_fp16 = add(x = matrix_ac_41_cast_fp16, y = matrix_bd_83_cast_fp16)[name = string("op_4797_cast_fp16")];
+            fp16 _inversed_scores_81_y_0_to_fp16 = const()[name = string("_inversed_scores_81_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_81_cast_fp16 = mul(x = var_4797_cast_fp16, y = _inversed_scores_81_y_0_to_fp16)[name = string("_inversed_scores_81_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_83_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_81_cast_fp16, cond = mask_11)[name = string("scores_83_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_4803_cast_fp16 = softmax(axis = var_58, x = scores_83_cast_fp16)[name = string("op_4803_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_1081_cast_fp16 = select(a = var_43_to_fp16, b = var_4803_cast_fp16, cond = mask_11)[name = string("input_1081_cast_fp16")];
+            bool x_533_transpose_x_0 = const()[name = string("x_533_transpose_x_0"), val = bool(false)];
+            bool x_533_transpose_y_0 = const()[name = string("x_533_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = v_41_cast_fp16)[name = string("transpose_179")];
+            tensor<fp16, [1, 8, 56, 128]> x_533_cast_fp16 = matmul(transpose_x = x_533_transpose_x_0, transpose_y = x_533_transpose_y_0, x = input_1081_cast_fp16, y = value_49_cast_fp16)[name = string("x_533_cast_fp16")];
+            tensor<int32, [4]> var_4807_perm_0 = const()[name = string("op_4807_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4808 = const()[name = string("op_4808"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_4807_cast_fp16 = transpose(perm = var_4807_perm_0, x = x_533_cast_fp16)[name = string("transpose_178")];
+            tensor<fp16, [1, 56, 1024]> input_1083_cast_fp16 = reshape(shape = var_4808, x = var_4807_cast_fp16)[name = string("input_1083_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(420536960))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421585600))))[name = string("encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421587712)))];
+            tensor<fp16, [1, 56, 1024]> linear_187_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized, x = input_1083_cast_fp16)[name = string("linear_187_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1087_cast_fp16 = add(x = input_1077_cast_fp16, y = linear_187_cast_fp16)[name = string("input_1087_cast_fp16")];
+            tensor<int32, [1]> x_537_axes_0 = const()[name = string("x_537_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421589824)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421591936)))];
+            tensor<fp16, [1, 56, 1024]> x_537_cast_fp16 = layer_norm(axes = x_537_axes_0, beta = encoder_layers_20_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_20_norm_conv_weight_to_fp16, x = input_1087_cast_fp16)[name = string("x_537_cast_fp16")];
+            tensor<int32, [3]> input_1089_perm_0 = const()[name = string("input_1089_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1091_pad_type_0 = const()[name = string("input_1091_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1091_strides_0 = const()[name = string("input_1091_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1091_pad_0 = const()[name = string("input_1091_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1091_dilations_0 = const()[name = string("input_1091_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1091_groups_0 = const()[name = string("input_1091_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421594048))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423691264))))[name = string("encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_1089_cast_fp16 = transpose(perm = input_1089_perm_0, x = x_537_cast_fp16)[name = string("transpose_177")];
+            tensor<fp16, [1, 2048, 56]> input_1091_cast_fp16 = conv(dilations = input_1091_dilations_0, groups = input_1091_groups_0, pad = input_1091_pad_0, pad_type = input_1091_pad_type_0, strides = input_1091_strides_0, weight = encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1089_cast_fp16)[name = string("input_1091_cast_fp16")];
+            int32 x_539_split_num_splits_0 = const()[name = string("x_539_split_num_splits_0"), val = int32(2)];
+            int32 x_539_split_axis_0 = const()[name = string("x_539_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_539_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_539_split_cast_fp16_1 = split(axis = x_539_split_axis_0, num_splits = x_539_split_num_splits_0, x = input_1091_cast_fp16)[name = string("x_539_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_539_split_1_sigmoid_cast_fp16 = sigmoid(x = x_539_split_cast_fp16_1)[name = string("x_539_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_539_cast_fp16 = mul(x = x_539_split_cast_fp16_0, y = x_539_split_1_sigmoid_cast_fp16)[name = string("x_539_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_1093_cast_fp16 = select(a = var_43_to_fp16, b = x_539_cast_fp16, cond = var_574)[name = string("input_1093_cast_fp16")];
+            bool new_x_83_interleave_0 = const()[name = string("new_x_83_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_83_cast_fp16 = concat(axis = var_58, interleave = new_x_83_interleave_0, values = (cache_83_cast_fp16, input_1093_cast_fp16))[name = string("new_x_83_cast_fp16")];
+            tensor<int32, [3]> var_4847_begin_0 = const()[name = string("op_4847_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_4847_end_0 = const()[name = string("op_4847_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_4847_end_mask_0 = const()[name = string("op_4847_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4847_cast_fp16 = slice_by_index(begin = var_4847_begin_0, end = var_4847_end_0, end_mask = var_4847_end_mask_0, x = new_x_83_cast_fp16)[name = string("op_4847_cast_fp16")];
+            string x_541_pad_type_0 = const()[name = string("x_541_pad_type_0"), val = string("valid")];
+            int32 x_541_groups_0 = const()[name = string("x_541_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_541_strides_0 = const()[name = string("x_541_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_541_pad_0 = const()[name = string("x_541_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_541_dilations_0 = const()[name = string("x_541_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423695424))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423704704))))[name = string("encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_541_cast_fp16 = conv(dilations = x_541_dilations_0, groups = x_541_groups_0, pad = x_541_pad_0, pad_type = x_541_pad_type_0, strides = x_541_strides_0, weight = encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_83_cast_fp16)[name = string("x_541_cast_fp16")];
+            tensor<int32, [3]> input_1095_perm_0 = const()[name = string("input_1095_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_543_axes_0 = const()[name = string("x_543_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423706816)))];
+            tensor<fp16, [1024]> encoder_layers_20_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423708928)))];
+            tensor<fp16, [1, 56, 1024]> input_1095_cast_fp16 = transpose(perm = input_1095_perm_0, x = x_541_cast_fp16)[name = string("transpose_176")];
+            tensor<fp16, [1, 56, 1024]> x_543_cast_fp16 = layer_norm(axes = x_543_axes_0, beta = encoder_layers_20_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_20_conv_batch_norm_weight_to_fp16, x = input_1095_cast_fp16)[name = string("x_543_cast_fp16")];
+            tensor<int32, [3]> input_1097_perm_0 = const()[name = string("input_1097_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_1097_cast_fp16 = transpose(perm = input_1097_perm_0, x = x_543_cast_fp16)[name = string("transpose_175")];
+            tensor<fp16, [1, 1024, 56]> input_1099_cast_fp16 = silu(x = input_1097_cast_fp16)[name = string("input_1099_cast_fp16")];
+            string x_545_pad_type_0 = const()[name = string("x_545_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_545_strides_0 = const()[name = string("x_545_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_545_pad_0 = const()[name = string("x_545_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_545_dilations_0 = const()[name = string("x_545_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_545_groups_0 = const()[name = string("x_545_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423711040))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424759680))))[name = string("encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_545_cast_fp16 = conv(dilations = x_545_dilations_0, groups = x_545_groups_0, pad = x_545_pad_0, pad_type = x_545_pad_type_0, strides = x_545_strides_0, weight = encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1099_cast_fp16)[name = string("x_545_cast_fp16")];
+            tensor<int32, [3]> input_1101_perm_0 = const()[name = string("input_1101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_1101_cast_fp16 = transpose(perm = input_1101_perm_0, x = x_545_cast_fp16)[name = string("transpose_174")];
+            tensor<fp16, [1, 56, 1024]> input_1103_cast_fp16 = add(x = input_1087_cast_fp16, y = input_1101_cast_fp16)[name = string("input_1103_cast_fp16")];
+            tensor<int32, [1]> input_1105_axes_0 = const()[name = string("input_1105_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424761792)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424763904)))];
+            tensor<fp16, [1, 56, 1024]> input_1105_cast_fp16 = layer_norm(axes = input_1105_axes_0, beta = encoder_layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_20_norm_feed_forward2_weight_to_fp16, x = input_1103_cast_fp16)[name = string("input_1105_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424766016))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428960384))))[name = string("encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_20_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428968640)))];
+            tensor<fp16, [1, 56, 4096]> linear_188_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1105_cast_fp16)[name = string("linear_188_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1109_cast_fp16 = silu(x = linear_188_cast_fp16)[name = string("input_1109_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428976896))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433171264))))[name = string("encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433173376)))];
+            tensor<fp16, [1, 56, 1024]> linear_189_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1109_cast_fp16)[name = string("linear_189_cast_fp16")];
+            fp16 var_4890_to_fp16 = const()[name = string("op_4890_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4891_cast_fp16 = mul(x = linear_189_cast_fp16, y = var_4890_to_fp16)[name = string("op_4891_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1115_cast_fp16 = add(x = input_1103_cast_fp16, y = var_4891_cast_fp16)[name = string("input_1115_cast_fp16")];
+            tensor<int32, [1]> input_1117_axes_0 = const()[name = string("input_1117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433175488)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433177600)))];
+            tensor<fp16, [1, 56, 1024]> input_1117_cast_fp16 = layer_norm(axes = input_1117_axes_0, beta = encoder_layers_20_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_20_norm_out_weight_to_fp16, x = input_1115_cast_fp16)[name = string("input_1117_cast_fp16")];
+            tensor<int32, [4]> cache_85_begin_0 = const()[name = string("cache_85_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
+            tensor<int32, [4]> cache_85_end_0 = const()[name = string("cache_85_end_0"), val = tensor<int32, [4]>([22, 1, 42, 1024])];
+            tensor<bool, [4]> cache_85_end_mask_0 = const()[name = string("cache_85_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_85_squeeze_mask_0 = const()[name = string("cache_85_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_85_cast_fp16 = slice_by_index(begin = cache_85_begin_0, end = cache_85_end_0, end_mask = cache_85_end_mask_0, squeeze_mask = cache_85_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_85_cast_fp16")];
+            tensor<int32, [4]> cache_87_begin_0 = const()[name = string("cache_87_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
+            tensor<int32, [4]> cache_87_end_0 = const()[name = string("cache_87_end_0"), val = tensor<int32, [4]>([22, 1, 1024, 8])];
+            tensor<bool, [4]> cache_87_end_mask_0 = const()[name = string("cache_87_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_87_squeeze_mask_0 = const()[name = string("cache_87_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_87_cast_fp16 = slice_by_index(begin = cache_87_begin_0, end = cache_87_end_0, end_mask = cache_87_end_mask_0, squeeze_mask = cache_87_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_87_cast_fp16")];
+            tensor<int32, [1]> input_1119_axes_0 = const()[name = string("input_1119_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433179712)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433181824)))];
+            tensor<fp16, [1, 56, 1024]> input_1119_cast_fp16 = layer_norm(axes = input_1119_axes_0, beta = encoder_layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_21_norm_feed_forward1_weight_to_fp16, x = input_1117_cast_fp16)[name = string("input_1119_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433183936))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437378304))))[name = string("encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_21_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437386560)))];
+            tensor<fp16, [1, 56, 4096]> linear_190_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1119_cast_fp16)[name = string("linear_190_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1123_cast_fp16 = silu(x = linear_190_cast_fp16)[name = string("input_1123_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437394816))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441589184))))[name = string("encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441591296)))];
+            tensor<fp16, [1, 56, 1024]> linear_191_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1123_cast_fp16)[name = string("linear_191_cast_fp16")];
+            fp16 var_4927_to_fp16 = const()[name = string("op_4927_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_4928_cast_fp16 = mul(x = linear_191_cast_fp16, y = var_4927_to_fp16)[name = string("op_4928_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1129_cast_fp16 = add(x = input_1117_cast_fp16, y = var_4928_cast_fp16)[name = string("input_1129_cast_fp16")];
+            tensor<int32, [1]> key_43_axes_0 = const()[name = string("key_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441593408)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441595520)))];
+            tensor<fp16, [1, 56, 1024]> key_43_cast_fp16 = layer_norm(axes = key_43_axes_0, beta = encoder_layers_21_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_21_norm_self_att_weight_to_fp16, x = input_1129_cast_fp16)[name = string("key_43_cast_fp16")];
+            bool input_1131_interleave_0 = const()[name = string("input_1131_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_1131_cast_fp16 = concat(axis = var_67, interleave = input_1131_interleave_0, values = (cache_85_cast_fp16, key_43_cast_fp16))[name = string("input_1131_cast_fp16")];
+            bool var_4956_interleave_0 = const()[name = string("op_4956_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_4956_cast_fp16 = concat(axis = var_67, interleave = var_4956_interleave_0, values = key_43_cast_fp16)[name = string("op_4956_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441597632))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442646272))))[name = string("encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442648384)))];
+            tensor<fp16, [1, 56, 1024]> linear_192_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized, x = key_43_cast_fp16)[name = string("linear_192_cast_fp16")];
+            tensor<int32, [4]> var_4961 = const()[name = string("op_4961"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_127_cast_fp16 = reshape(shape = var_4961, x = linear_192_cast_fp16)[name = string("q_127_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442650496))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443699136))))[name = string("encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443701248)))];
+            tensor<fp16, [1, 98, 1024]> linear_193_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_193_cast_fp16")];
+            tensor<int32, [4]> var_4966 = const()[name = string("op_4966"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_85_cast_fp16 = reshape(shape = var_4966, x = linear_193_cast_fp16)[name = string("k_85_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443703360))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444752000))))[name = string("encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444754112)))];
+            tensor<fp16, [1, 98, 1024]> linear_194_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_194_cast_fp16")];
+            tensor<int32, [4]> var_4971 = const()[name = string("op_4971"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_43_cast_fp16 = reshape(shape = var_4971, x = linear_194_cast_fp16)[name = string("v_43_cast_fp16")];
+            tensor<int32, [4]> value_51_perm_0 = const()[name = string("value_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_21_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444756224)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4984_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_u_to_fp16)[name = string("op_4984_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_21_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444758336)))];
+            tensor<fp16, [1, 56, 8, 128]> var_4986_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_v_to_fp16)[name = string("op_4986_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_43_perm_0 = const()[name = string("q_with_bias_v_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_553_transpose_x_0 = const()[name = string("x_553_transpose_x_0"), val = bool(false)];
+            bool x_553_transpose_y_0 = const()[name = string("x_553_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_4988_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444760448))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444960192))))[name = string("op_4988_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_43_cast_fp16 = transpose(perm = q_with_bias_v_43_perm_0, x = var_4986_cast_fp16)[name = string("transpose_173")];
+            tensor<fp16, [1, 8, 56, 195]> x_553_cast_fp16 = matmul(transpose_x = x_553_transpose_x_0, transpose_y = x_553_transpose_y_0, x = q_with_bias_v_43_cast_fp16, y = op_4988_to_fp16_quantized)[name = string("x_553_cast_fp16")];
+            tensor<int32, [8]> x_555_pad_0 = const()[name = string("x_555_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_555_mode_0 = const()[name = string("x_555_mode_0"), val = string("constant")];
+            fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_555_cast_fp16 = pad(constant_val = const_352_to_fp16, mode = x_555_mode_0, pad = x_555_pad_0, x = x_553_cast_fp16)[name = string("x_555_cast_fp16")];
+            tensor<int32, [4]> var_4996 = const()[name = string("op_4996"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_557_cast_fp16 = reshape(shape = var_4996, x = x_555_cast_fp16)[name = string("x_557_cast_fp16")];
+            tensor<int32, [4]> var_5000_begin_0 = const()[name = string("op_5000_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5000_end_0 = const()[name = string("op_5000_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_5000_end_mask_0 = const()[name = string("op_5000_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_5000_cast_fp16 = slice_by_index(begin = var_5000_begin_0, end = var_5000_end_0, end_mask = var_5000_end_mask_0, x = x_557_cast_fp16)[name = string("op_5000_cast_fp16")];
+            tensor<int32, [4]> var_5001 = const()[name = string("op_5001"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_85_cast_fp16 = reshape(shape = var_5001, x = var_5000_cast_fp16)[name = string("matrix_bd_85_cast_fp16")];
+            bool matrix_ac_43_transpose_x_0 = const()[name = string("matrix_ac_43_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_43_transpose_y_0 = const()[name = string("matrix_ac_43_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_138_perm_0 = const()[name = string("transpose_138_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_139_perm_0 = const()[name = string("transpose_139_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_139 = transpose(perm = transpose_139_perm_0, x = k_85_cast_fp16)[name = string("transpose_171")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_138 = transpose(perm = transpose_138_perm_0, x = var_4984_cast_fp16)[name = string("transpose_172")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_43_cast_fp16 = matmul(transpose_x = matrix_ac_43_transpose_x_0, transpose_y = matrix_ac_43_transpose_y_0, x = transpose_138, y = transpose_139)[name = string("matrix_ac_43_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_87_begin_0 = const()[name = string("matrix_bd_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_87_end_0 = const()[name = string("matrix_bd_87_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_87_end_mask_0 = const()[name = string("matrix_bd_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_87_cast_fp16 = slice_by_index(begin = matrix_bd_87_begin_0, end = matrix_bd_87_end_0, end_mask = matrix_bd_87_end_mask_0, x = matrix_bd_85_cast_fp16)[name = string("matrix_bd_87_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_5010_cast_fp16 = add(x = matrix_ac_43_cast_fp16, y = matrix_bd_87_cast_fp16)[name = string("op_5010_cast_fp16")];
+            fp16 _inversed_scores_85_y_0_to_fp16 = const()[name = string("_inversed_scores_85_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_85_cast_fp16 = mul(x = var_5010_cast_fp16, y = _inversed_scores_85_y_0_to_fp16)[name = string("_inversed_scores_85_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_87_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_85_cast_fp16, cond = mask_11)[name = string("scores_87_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_5016_cast_fp16 = softmax(axis = var_58, x = scores_87_cast_fp16)[name = string("op_5016_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_1133_cast_fp16 = select(a = var_43_to_fp16, b = var_5016_cast_fp16, cond = mask_11)[name = string("input_1133_cast_fp16")];
+            bool x_559_transpose_x_0 = const()[name = string("x_559_transpose_x_0"), val = bool(false)];
+            bool x_559_transpose_y_0 = const()[name = string("x_559_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = v_43_cast_fp16)[name = string("transpose_170")];
+            tensor<fp16, [1, 8, 56, 128]> x_559_cast_fp16 = matmul(transpose_x = x_559_transpose_x_0, transpose_y = x_559_transpose_y_0, x = input_1133_cast_fp16, y = value_51_cast_fp16)[name = string("x_559_cast_fp16")];
+            tensor<int32, [4]> var_5020_perm_0 = const()[name = string("op_5020_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5021 = const()[name = string("op_5021"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_5020_cast_fp16 = transpose(perm = var_5020_perm_0, x = x_559_cast_fp16)[name = string("transpose_169")];
+            tensor<fp16, [1, 56, 1024]> input_1135_cast_fp16 = reshape(shape = var_5021, x = var_5020_cast_fp16)[name = string("input_1135_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444960704)))];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447057920)))];
+            tensor<fp16, [1, 56, 1024]> linear_196_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_out_weight_to_fp16, x = input_1135_cast_fp16)[name = string("linear_196_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1139_cast_fp16 = add(x = input_1129_cast_fp16, y = linear_196_cast_fp16)[name = string("input_1139_cast_fp16")];
+            tensor<int32, [1]> x_563_axes_0 = const()[name = string("x_563_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447060032)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447062144)))];
+            tensor<fp16, [1, 56, 1024]> x_563_cast_fp16 = layer_norm(axes = x_563_axes_0, beta = encoder_layers_21_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_21_norm_conv_weight_to_fp16, x = input_1139_cast_fp16)[name = string("x_563_cast_fp16")];
+            tensor<int32, [3]> input_1141_perm_0 = const()[name = string("input_1141_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1143_pad_type_0 = const()[name = string("input_1143_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1143_strides_0 = const()[name = string("input_1143_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1143_pad_0 = const()[name = string("input_1143_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1143_dilations_0 = const()[name = string("input_1143_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1143_groups_0 = const()[name = string("input_1143_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(447064256))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449161472))))[name = string("encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_1141_cast_fp16 = transpose(perm = input_1141_perm_0, x = x_563_cast_fp16)[name = string("transpose_168")];
+            tensor<fp16, [1, 2048, 56]> input_1143_cast_fp16 = conv(dilations = input_1143_dilations_0, groups = input_1143_groups_0, pad = input_1143_pad_0, pad_type = input_1143_pad_type_0, strides = input_1143_strides_0, weight = encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1141_cast_fp16)[name = string("input_1143_cast_fp16")];
+            int32 x_565_split_num_splits_0 = const()[name = string("x_565_split_num_splits_0"), val = int32(2)];
+            int32 x_565_split_axis_0 = const()[name = string("x_565_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_565_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_565_split_cast_fp16_1 = split(axis = x_565_split_axis_0, num_splits = x_565_split_num_splits_0, x = input_1143_cast_fp16)[name = string("x_565_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_565_split_1_sigmoid_cast_fp16 = sigmoid(x = x_565_split_cast_fp16_1)[name = string("x_565_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_565_cast_fp16 = mul(x = x_565_split_cast_fp16_0, y = x_565_split_1_sigmoid_cast_fp16)[name = string("x_565_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_1145_cast_fp16 = select(a = var_43_to_fp16, b = x_565_cast_fp16, cond = var_574)[name = string("input_1145_cast_fp16")];
+            bool new_x_87_interleave_0 = const()[name = string("new_x_87_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_87_cast_fp16 = concat(axis = var_58, interleave = new_x_87_interleave_0, values = (cache_87_cast_fp16, input_1145_cast_fp16))[name = string("new_x_87_cast_fp16")];
+            tensor<int32, [3]> var_5060_begin_0 = const()[name = string("op_5060_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_5060_end_0 = const()[name = string("op_5060_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_5060_end_mask_0 = const()[name = string("op_5060_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_5060_cast_fp16 = slice_by_index(begin = var_5060_begin_0, end = var_5060_end_0, end_mask = var_5060_end_mask_0, x = new_x_87_cast_fp16)[name = string("op_5060_cast_fp16")];
+            string x_567_pad_type_0 = const()[name = string("x_567_pad_type_0"), val = string("valid")];
+            int32 x_567_groups_0 = const()[name = string("x_567_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_567_strides_0 = const()[name = string("x_567_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_567_pad_0 = const()[name = string("x_567_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_567_dilations_0 = const()[name = string("x_567_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449165632))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449174912))))[name = string("encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_567_cast_fp16 = conv(dilations = x_567_dilations_0, groups = x_567_groups_0, pad = x_567_pad_0, pad_type = x_567_pad_type_0, strides = x_567_strides_0, weight = encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_87_cast_fp16)[name = string("x_567_cast_fp16")];
+            tensor<int32, [3]> input_1147_perm_0 = const()[name = string("input_1147_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_569_axes_0 = const()[name = string("x_569_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449177024)))];
+            tensor<fp16, [1024]> encoder_layers_21_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449179136)))];
+            tensor<fp16, [1, 56, 1024]> input_1147_cast_fp16 = transpose(perm = input_1147_perm_0, x = x_567_cast_fp16)[name = string("transpose_167")];
+            tensor<fp16, [1, 56, 1024]> x_569_cast_fp16 = layer_norm(axes = x_569_axes_0, beta = encoder_layers_21_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_21_conv_batch_norm_weight_to_fp16, x = input_1147_cast_fp16)[name = string("x_569_cast_fp16")];
+            tensor<int32, [3]> input_1149_perm_0 = const()[name = string("input_1149_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_1149_cast_fp16 = transpose(perm = input_1149_perm_0, x = x_569_cast_fp16)[name = string("transpose_166")];
+            tensor<fp16, [1, 1024, 56]> input_1151_cast_fp16 = silu(x = input_1149_cast_fp16)[name = string("input_1151_cast_fp16")];
+            string x_571_pad_type_0 = const()[name = string("x_571_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_571_strides_0 = const()[name = string("x_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_571_pad_0 = const()[name = string("x_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_571_dilations_0 = const()[name = string("x_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_571_groups_0 = const()[name = string("x_571_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449181248))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450229888))))[name = string("encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_571_cast_fp16 = conv(dilations = x_571_dilations_0, groups = x_571_groups_0, pad = x_571_pad_0, pad_type = x_571_pad_type_0, strides = x_571_strides_0, weight = encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1151_cast_fp16)[name = string("x_571_cast_fp16")];
+            tensor<int32, [3]> input_1153_perm_0 = const()[name = string("input_1153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_1153_cast_fp16 = transpose(perm = input_1153_perm_0, x = x_571_cast_fp16)[name = string("transpose_165")];
+            tensor<fp16, [1, 56, 1024]> input_1155_cast_fp16 = add(x = input_1139_cast_fp16, y = input_1153_cast_fp16)[name = string("input_1155_cast_fp16")];
+            tensor<int32, [1]> input_1157_axes_0 = const()[name = string("input_1157_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450232000)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450234112)))];
+            tensor<fp16, [1, 56, 1024]> input_1157_cast_fp16 = layer_norm(axes = input_1157_axes_0, beta = encoder_layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_21_norm_feed_forward2_weight_to_fp16, x = input_1155_cast_fp16)[name = string("input_1157_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_21_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(450236224)))];
+            tensor<fp16, [4096]> encoder_layers_21_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458624896)))];
+            tensor<fp16, [1, 56, 4096]> linear_197_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear1_weight_to_fp16, x = input_1157_cast_fp16)[name = string("linear_197_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1161_cast_fp16 = silu(x = linear_197_cast_fp16)[name = string("input_1161_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_21_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458633152)))];
+            tensor<fp16, [1024]> encoder_layers_21_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467021824)))];
+            tensor<fp16, [1, 56, 1024]> linear_198_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear2_weight_to_fp16, x = input_1161_cast_fp16)[name = string("linear_198_cast_fp16")];
+            fp16 var_5103_to_fp16 = const()[name = string("op_5103_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_5104_cast_fp16 = mul(x = linear_198_cast_fp16, y = var_5103_to_fp16)[name = string("op_5104_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1167_cast_fp16 = add(x = input_1155_cast_fp16, y = var_5104_cast_fp16)[name = string("input_1167_cast_fp16")];
+            tensor<int32, [1]> input_1169_axes_0 = const()[name = string("input_1169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467023936)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467026048)))];
+            tensor<fp16, [1, 56, 1024]> input_1169_cast_fp16 = layer_norm(axes = input_1169_axes_0, beta = encoder_layers_21_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_21_norm_out_weight_to_fp16, x = input_1167_cast_fp16)[name = string("input_1169_cast_fp16")];
+            tensor<int32, [4]> cache_89_begin_0 = const()[name = string("cache_89_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
+            tensor<int32, [4]> cache_89_end_0 = const()[name = string("cache_89_end_0"), val = tensor<int32, [4]>([23, 1, 42, 1024])];
+            tensor<bool, [4]> cache_89_end_mask_0 = const()[name = string("cache_89_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_89_squeeze_mask_0 = const()[name = string("cache_89_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_89_cast_fp16 = slice_by_index(begin = cache_89_begin_0, end = cache_89_end_0, end_mask = cache_89_end_mask_0, squeeze_mask = cache_89_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_89_cast_fp16")];
+            tensor<int32, [4]> cache_91_begin_0 = const()[name = string("cache_91_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
+            tensor<int32, [4]> cache_91_end_0 = const()[name = string("cache_91_end_0"), val = tensor<int32, [4]>([23, 1, 1024, 8])];
+            tensor<bool, [4]> cache_91_end_mask_0 = const()[name = string("cache_91_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_91_squeeze_mask_0 = const()[name = string("cache_91_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_91_cast_fp16 = slice_by_index(begin = cache_91_begin_0, end = cache_91_end_0, end_mask = cache_91_end_mask_0, squeeze_mask = cache_91_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_91_cast_fp16")];
+            tensor<int32, [1]> input_1171_axes_0 = const()[name = string("input_1171_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467028160)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467030272)))];
+            tensor<fp16, [1, 56, 1024]> input_1171_cast_fp16 = layer_norm(axes = input_1171_axes_0, beta = encoder_layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_22_norm_feed_forward1_weight_to_fp16, x = input_1169_cast_fp16)[name = string("input_1171_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_22_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467032384)))];
+            tensor<fp16, [4096]> encoder_layers_22_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475421056)))];
+            tensor<fp16, [1, 56, 4096]> linear_199_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear1_weight_to_fp16, x = input_1171_cast_fp16)[name = string("linear_199_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1175_cast_fp16 = silu(x = linear_199_cast_fp16)[name = string("input_1175_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_22_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475429312)))];
+            tensor<fp16, [1024]> encoder_layers_22_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483817984)))];
+            tensor<fp16, [1, 56, 1024]> linear_200_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear2_weight_to_fp16, x = input_1175_cast_fp16)[name = string("linear_200_cast_fp16")];
+            fp16 var_5140_to_fp16 = const()[name = string("op_5140_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_5141_cast_fp16 = mul(x = linear_200_cast_fp16, y = var_5140_to_fp16)[name = string("op_5141_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1181_cast_fp16 = add(x = input_1169_cast_fp16, y = var_5141_cast_fp16)[name = string("input_1181_cast_fp16")];
+            tensor<int32, [1]> key_45_axes_0 = const()[name = string("key_45_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483820096)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483822208)))];
+            tensor<fp16, [1, 56, 1024]> key_45_cast_fp16 = layer_norm(axes = key_45_axes_0, beta = encoder_layers_22_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_22_norm_self_att_weight_to_fp16, x = input_1181_cast_fp16)[name = string("key_45_cast_fp16")];
+            bool input_1183_interleave_0 = const()[name = string("input_1183_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_1183_cast_fp16 = concat(axis = var_67, interleave = input_1183_interleave_0, values = (cache_89_cast_fp16, key_45_cast_fp16))[name = string("input_1183_cast_fp16")];
+            bool var_5169_interleave_0 = const()[name = string("op_5169_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> var_5169_cast_fp16 = concat(axis = var_67, interleave = var_5169_interleave_0, values = key_45_cast_fp16)[name = string("op_5169_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483824320)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485921536)))];
+            tensor<fp16, [1, 56, 1024]> linear_201_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_q_weight_to_fp16, x = key_45_cast_fp16)[name = string("linear_201_cast_fp16")];
+            tensor<int32, [4]> var_5174 = const()[name = string("op_5174"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_133_cast_fp16 = reshape(shape = var_5174, x = linear_201_cast_fp16)[name = string("q_133_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485923648)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488020864)))];
+            tensor<fp16, [1, 98, 1024]> linear_202_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_k_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_202_cast_fp16")];
+            tensor<int32, [4]> var_5179 = const()[name = string("op_5179"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_89_cast_fp16 = reshape(shape = var_5179, x = linear_202_cast_fp16)[name = string("k_89_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488022976)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490120192)))];
+            tensor<fp16, [1, 98, 1024]> linear_203_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_v_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_203_cast_fp16")];
+            tensor<int32, [4]> var_5184 = const()[name = string("op_5184"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_45_cast_fp16 = reshape(shape = var_5184, x = linear_203_cast_fp16)[name = string("v_45_cast_fp16")];
+            tensor<int32, [4]> value_53_perm_0 = const()[name = string("value_53_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_22_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490122304)))];
+            tensor<fp16, [1, 56, 8, 128]> var_5197_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_u_to_fp16)[name = string("op_5197_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_22_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490124416)))];
+            tensor<fp16, [1, 56, 8, 128]> var_5199_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_v_to_fp16)[name = string("op_5199_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_45_perm_0 = const()[name = string("q_with_bias_v_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_579_transpose_x_0 = const()[name = string("x_579_transpose_x_0"), val = bool(false)];
+            bool x_579_transpose_y_0 = const()[name = string("x_579_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_5201_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490126528))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490326272))))[name = string("op_5201_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_45_cast_fp16 = transpose(perm = q_with_bias_v_45_perm_0, x = var_5199_cast_fp16)[name = string("transpose_164")];
+            tensor<fp16, [1, 8, 56, 195]> x_579_cast_fp16 = matmul(transpose_x = x_579_transpose_x_0, transpose_y = x_579_transpose_y_0, x = q_with_bias_v_45_cast_fp16, y = op_5201_to_fp16_quantized)[name = string("x_579_cast_fp16")];
+            tensor<int32, [8]> x_581_pad_0 = const()[name = string("x_581_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_581_mode_0 = const()[name = string("x_581_mode_0"), val = string("constant")];
+            fp16 const_365_to_fp16 = const()[name = string("const_365_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_581_cast_fp16 = pad(constant_val = const_365_to_fp16, mode = x_581_mode_0, pad = x_581_pad_0, x = x_579_cast_fp16)[name = string("x_581_cast_fp16")];
+            tensor<int32, [4]> var_5209 = const()[name = string("op_5209"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_583_cast_fp16 = reshape(shape = var_5209, x = x_581_cast_fp16)[name = string("x_583_cast_fp16")];
+            tensor<int32, [4]> var_5213_begin_0 = const()[name = string("op_5213_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5213_end_0 = const()[name = string("op_5213_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_5213_end_mask_0 = const()[name = string("op_5213_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_5213_cast_fp16 = slice_by_index(begin = var_5213_begin_0, end = var_5213_end_0, end_mask = var_5213_end_mask_0, x = x_583_cast_fp16)[name = string("op_5213_cast_fp16")];
+            tensor<int32, [4]> var_5214 = const()[name = string("op_5214"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_89_cast_fp16 = reshape(shape = var_5214, x = var_5213_cast_fp16)[name = string("matrix_bd_89_cast_fp16")];
+            bool matrix_ac_45_transpose_x_0 = const()[name = string("matrix_ac_45_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_45_transpose_y_0 = const()[name = string("matrix_ac_45_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_140_perm_0 = const()[name = string("transpose_140_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_141_perm_0 = const()[name = string("transpose_141_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_141 = transpose(perm = transpose_141_perm_0, x = k_89_cast_fp16)[name = string("transpose_162")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_140 = transpose(perm = transpose_140_perm_0, x = var_5197_cast_fp16)[name = string("transpose_163")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_45_cast_fp16 = matmul(transpose_x = matrix_ac_45_transpose_x_0, transpose_y = matrix_ac_45_transpose_y_0, x = transpose_140, y = transpose_141)[name = string("matrix_ac_45_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_91_begin_0 = const()[name = string("matrix_bd_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_91_end_0 = const()[name = string("matrix_bd_91_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_91_end_mask_0 = const()[name = string("matrix_bd_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_91_cast_fp16 = slice_by_index(begin = matrix_bd_91_begin_0, end = matrix_bd_91_end_0, end_mask = matrix_bd_91_end_mask_0, x = matrix_bd_89_cast_fp16)[name = string("matrix_bd_91_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_5223_cast_fp16 = add(x = matrix_ac_45_cast_fp16, y = matrix_bd_91_cast_fp16)[name = string("op_5223_cast_fp16")];
+            fp16 _inversed_scores_89_y_0_to_fp16 = const()[name = string("_inversed_scores_89_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_89_cast_fp16 = mul(x = var_5223_cast_fp16, y = _inversed_scores_89_y_0_to_fp16)[name = string("_inversed_scores_89_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_91_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_89_cast_fp16, cond = mask_11)[name = string("scores_91_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_5229_cast_fp16 = softmax(axis = var_58, x = scores_91_cast_fp16)[name = string("op_5229_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_1185_cast_fp16 = select(a = var_43_to_fp16, b = var_5229_cast_fp16, cond = mask_11)[name = string("input_1185_cast_fp16")];
+            bool x_585_transpose_x_0 = const()[name = string("x_585_transpose_x_0"), val = bool(false)];
+            bool x_585_transpose_y_0 = const()[name = string("x_585_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = v_45_cast_fp16)[name = string("transpose_161")];
+            tensor<fp16, [1, 8, 56, 128]> x_585_cast_fp16 = matmul(transpose_x = x_585_transpose_x_0, transpose_y = x_585_transpose_y_0, x = input_1185_cast_fp16, y = value_53_cast_fp16)[name = string("x_585_cast_fp16")];
+            tensor<int32, [4]> var_5233_perm_0 = const()[name = string("op_5233_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5234 = const()[name = string("op_5234"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_5233_cast_fp16 = transpose(perm = var_5233_perm_0, x = x_585_cast_fp16)[name = string("transpose_160")];
+            tensor<fp16, [1, 56, 1024]> input_1187_cast_fp16 = reshape(shape = var_5234, x = var_5233_cast_fp16)[name = string("input_1187_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490326784)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492424000)))];
+            tensor<fp16, [1, 56, 1024]> linear_205_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_out_weight_to_fp16, x = input_1187_cast_fp16)[name = string("linear_205_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1191_cast_fp16 = add(x = input_1181_cast_fp16, y = linear_205_cast_fp16)[name = string("input_1191_cast_fp16")];
+            tensor<int32, [1]> x_589_axes_0 = const()[name = string("x_589_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492426112)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492428224)))];
+            tensor<fp16, [1, 56, 1024]> x_589_cast_fp16 = layer_norm(axes = x_589_axes_0, beta = encoder_layers_22_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_22_norm_conv_weight_to_fp16, x = input_1191_cast_fp16)[name = string("x_589_cast_fp16")];
+            tensor<int32, [3]> input_1193_perm_0 = const()[name = string("input_1193_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1195_pad_type_0 = const()[name = string("input_1195_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1195_strides_0 = const()[name = string("input_1195_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1195_pad_0 = const()[name = string("input_1195_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1195_dilations_0 = const()[name = string("input_1195_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1195_groups_0 = const()[name = string("input_1195_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492430336))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494527552))))[name = string("encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_1193_cast_fp16 = transpose(perm = input_1193_perm_0, x = x_589_cast_fp16)[name = string("transpose_159")];
+            tensor<fp16, [1, 2048, 56]> input_1195_cast_fp16 = conv(dilations = input_1195_dilations_0, groups = input_1195_groups_0, pad = input_1195_pad_0, pad_type = input_1195_pad_type_0, strides = input_1195_strides_0, weight = encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1193_cast_fp16)[name = string("input_1195_cast_fp16")];
+            int32 x_591_split_num_splits_0 = const()[name = string("x_591_split_num_splits_0"), val = int32(2)];
+            int32 x_591_split_axis_0 = const()[name = string("x_591_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_591_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_591_split_cast_fp16_1 = split(axis = x_591_split_axis_0, num_splits = x_591_split_num_splits_0, x = input_1195_cast_fp16)[name = string("x_591_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_591_split_1_sigmoid_cast_fp16 = sigmoid(x = x_591_split_cast_fp16_1)[name = string("x_591_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_591_cast_fp16 = mul(x = x_591_split_cast_fp16_0, y = x_591_split_1_sigmoid_cast_fp16)[name = string("x_591_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_1197_cast_fp16 = select(a = var_43_to_fp16, b = x_591_cast_fp16, cond = var_574)[name = string("input_1197_cast_fp16")];
+            bool new_x_91_interleave_0 = const()[name = string("new_x_91_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_91_cast_fp16 = concat(axis = var_58, interleave = new_x_91_interleave_0, values = (cache_91_cast_fp16, input_1197_cast_fp16))[name = string("new_x_91_cast_fp16")];
+            tensor<int32, [3]> var_5273_begin_0 = const()[name = string("op_5273_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> var_5273_end_0 = const()[name = string("op_5273_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> var_5273_end_mask_0 = const()[name = string("op_5273_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_5273_cast_fp16 = slice_by_index(begin = var_5273_begin_0, end = var_5273_end_0, end_mask = var_5273_end_mask_0, x = new_x_91_cast_fp16)[name = string("op_5273_cast_fp16")];
+            string x_593_pad_type_0 = const()[name = string("x_593_pad_type_0"), val = string("valid")];
+            int32 x_593_groups_0 = const()[name = string("x_593_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_593_strides_0 = const()[name = string("x_593_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_593_pad_0 = const()[name = string("x_593_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_593_dilations_0 = const()[name = string("x_593_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494531712))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494540992))))[name = string("encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_593_cast_fp16 = conv(dilations = x_593_dilations_0, groups = x_593_groups_0, pad = x_593_pad_0, pad_type = x_593_pad_type_0, strides = x_593_strides_0, weight = encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_91_cast_fp16)[name = string("x_593_cast_fp16")];
+            tensor<int32, [3]> input_1199_perm_0 = const()[name = string("input_1199_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_595_axes_0 = const()[name = string("x_595_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494543104)))];
+            tensor<fp16, [1024]> encoder_layers_22_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494545216)))];
+            tensor<fp16, [1, 56, 1024]> input_1199_cast_fp16 = transpose(perm = input_1199_perm_0, x = x_593_cast_fp16)[name = string("transpose_158")];
+            tensor<fp16, [1, 56, 1024]> x_595_cast_fp16 = layer_norm(axes = x_595_axes_0, beta = encoder_layers_22_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_22_conv_batch_norm_weight_to_fp16, x = input_1199_cast_fp16)[name = string("x_595_cast_fp16")];
+            tensor<int32, [3]> input_1201_perm_0 = const()[name = string("input_1201_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_1201_cast_fp16 = transpose(perm = input_1201_perm_0, x = x_595_cast_fp16)[name = string("transpose_157")];
+            tensor<fp16, [1, 1024, 56]> input_1203_cast_fp16 = silu(x = input_1201_cast_fp16)[name = string("input_1203_cast_fp16")];
+            string x_597_pad_type_0 = const()[name = string("x_597_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_597_strides_0 = const()[name = string("x_597_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_597_pad_0 = const()[name = string("x_597_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_597_dilations_0 = const()[name = string("x_597_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_597_groups_0 = const()[name = string("x_597_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494547328))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495595968))))[name = string("encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_597_cast_fp16 = conv(dilations = x_597_dilations_0, groups = x_597_groups_0, pad = x_597_pad_0, pad_type = x_597_pad_type_0, strides = x_597_strides_0, weight = encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1203_cast_fp16)[name = string("x_597_cast_fp16")];
+            tensor<int32, [3]> input_1205_perm_0 = const()[name = string("input_1205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_1205_cast_fp16 = transpose(perm = input_1205_perm_0, x = x_597_cast_fp16)[name = string("transpose_156")];
+            tensor<fp16, [1, 56, 1024]> input_1207_cast_fp16 = add(x = input_1191_cast_fp16, y = input_1205_cast_fp16)[name = string("input_1207_cast_fp16")];
+            tensor<int32, [1]> input_1209_axes_0 = const()[name = string("input_1209_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495598080)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495600192)))];
+            tensor<fp16, [1, 56, 1024]> input_1209_cast_fp16 = layer_norm(axes = input_1209_axes_0, beta = encoder_layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_22_norm_feed_forward2_weight_to_fp16, x = input_1207_cast_fp16)[name = string("input_1209_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_22_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495602304)))];
+            tensor<fp16, [4096]> encoder_layers_22_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503990976)))];
+            tensor<fp16, [1, 56, 4096]> linear_206_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear1_weight_to_fp16, x = input_1209_cast_fp16)[name = string("linear_206_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1213_cast_fp16 = silu(x = linear_206_cast_fp16)[name = string("input_1213_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_22_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503999232)))];
+            tensor<fp16, [1024]> encoder_layers_22_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512387904)))];
+            tensor<fp16, [1, 56, 1024]> linear_207_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear2_weight_to_fp16, x = input_1213_cast_fp16)[name = string("linear_207_cast_fp16")];
+            fp16 var_5316_to_fp16 = const()[name = string("op_5316_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_5317_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_5316_to_fp16)[name = string("op_5317_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1219_cast_fp16 = add(x = input_1207_cast_fp16, y = var_5317_cast_fp16)[name = string("input_1219_cast_fp16")];
+            tensor<int32, [1]> input_1221_axes_0 = const()[name = string("input_1221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512390016)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512392128)))];
+            tensor<fp16, [1, 56, 1024]> input_1221_cast_fp16 = layer_norm(axes = input_1221_axes_0, beta = encoder_layers_22_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_22_norm_out_weight_to_fp16, x = input_1219_cast_fp16)[name = string("input_1221_cast_fp16")];
+            tensor<int32, [4]> cache_93_begin_0 = const()[name = string("cache_93_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
+            tensor<int32, [4]> cache_93_end_0 = const()[name = string("cache_93_end_0"), val = tensor<int32, [4]>([24, 1, 42, 1024])];
+            tensor<bool, [4]> cache_93_end_mask_0 = const()[name = string("cache_93_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_93_squeeze_mask_0 = const()[name = string("cache_93_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_93_cast_fp16 = slice_by_index(begin = cache_93_begin_0, end = cache_93_end_0, end_mask = cache_93_end_mask_0, squeeze_mask = cache_93_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_93_cast_fp16")];
+            tensor<int32, [4]> cache_begin_0 = const()[name = string("cache_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
+            tensor<int32, [4]> cache_end_0 = const()[name = string("cache_end_0"), val = tensor<int32, [4]>([24, 1, 1024, 8])];
+            tensor<bool, [4]> cache_end_mask_0 = const()[name = string("cache_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_squeeze_mask_0 = const()[name = string("cache_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_cast_fp16 = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_cast_fp16")];
+            tensor<int32, [1]> input_1223_axes_0 = const()[name = string("input_1223_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512394240)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512396352)))];
+            tensor<fp16, [1, 56, 1024]> input_1223_cast_fp16 = layer_norm(axes = input_1223_axes_0, beta = encoder_layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_23_norm_feed_forward1_weight_to_fp16, x = input_1221_cast_fp16)[name = string("input_1223_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_23_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512398464)))];
+            tensor<fp16, [4096]> encoder_layers_23_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520787136)))];
+            tensor<fp16, [1, 56, 4096]> linear_208_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear1_weight_to_fp16, x = input_1223_cast_fp16)[name = string("linear_208_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1227_cast_fp16 = silu(x = linear_208_cast_fp16)[name = string("input_1227_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_23_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520795392)))];
+            tensor<fp16, [1024]> encoder_layers_23_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529184064)))];
+            tensor<fp16, [1, 56, 1024]> linear_209_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear2_weight_to_fp16, x = input_1227_cast_fp16)[name = string("linear_209_cast_fp16")];
+            fp16 var_5353_to_fp16 = const()[name = string("op_5353_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_5354_cast_fp16 = mul(x = linear_209_cast_fp16, y = var_5353_to_fp16)[name = string("op_5354_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1233_cast_fp16 = add(x = input_1221_cast_fp16, y = var_5354_cast_fp16)[name = string("input_1233_cast_fp16")];
+            tensor<int32, [1]> key_axes_0 = const()[name = string("key_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529186176)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529188288)))];
+            tensor<fp16, [1, 56, 1024]> key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_23_norm_self_att_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_23_norm_self_att_weight_to_fp16, x = input_1233_cast_fp16)[name = string("key_cast_fp16")];
+            bool input_1235_interleave_0 = const()[name = string("input_1235_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 98, 1024]> input_1235_cast_fp16 = concat(axis = var_67, interleave = input_1235_interleave_0, values = (cache_93_cast_fp16, key_cast_fp16))[name = string("input_1235_cast_fp16")];
+            bool cache_last_channel_cur_interleave_0 = const()[name = string("cache_last_channel_cur_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1024]> cache_last_channel_cur_cast_fp16 = concat(axis = var_67, interleave = cache_last_channel_cur_interleave_0, values = key_cast_fp16)[name = string("cache_last_channel_cur_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529190400)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531287616)))];
+            tensor<fp16, [1, 56, 1024]> linear_210_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_q_weight_to_fp16, x = key_cast_fp16)[name = string("linear_210_cast_fp16")];
+            tensor<int32, [4]> var_5387 = const()[name = string("op_5387"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 56, 8, 128]> q_139_cast_fp16 = reshape(shape = var_5387, x = linear_210_cast_fp16)[name = string("q_139_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531289728)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533386944)))];
+            tensor<fp16, [1, 98, 1024]> linear_211_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_k_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_211_cast_fp16")];
+            tensor<int32, [4]> var_5392 = const()[name = string("op_5392"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> k_93_cast_fp16 = reshape(shape = var_5392, x = linear_211_cast_fp16)[name = string("k_93_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533389056)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535486272)))];
+            tensor<fp16, [1, 98, 1024]> linear_212_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_v_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_212_cast_fp16")];
+            tensor<int32, [4]> var_5397 = const()[name = string("op_5397"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 98, 8, 128]> v_cast_fp16 = reshape(shape = var_5397, x = linear_212_cast_fp16)[name = string("v_cast_fp16")];
+            tensor<int32, [4]> value_perm_0 = const()[name = string("value_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_23_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535488384)))];
+            tensor<fp16, [1, 56, 8, 128]> var_5410_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_u_to_fp16)[name = string("op_5410_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_23_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535490496)))];
+            tensor<fp16, [1, 56, 8, 128]> var_5412_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_v_to_fp16)[name = string("op_5412_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_perm_0 = const()[name = string("q_with_bias_v_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_605_transpose_x_0 = const()[name = string("x_605_transpose_x_0"), val = bool(false)];
+            bool x_605_transpose_y_0 = const()[name = string("x_605_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 195]> op_5414_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535492608))), scale = tensor<fp16, [1, 1, 1, 195]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535692352))))[name = string("op_5414_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 56, 128]> q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_5412_cast_fp16)[name = string("transpose_155")];
+            tensor<fp16, [1, 8, 56, 195]> x_605_cast_fp16 = matmul(transpose_x = x_605_transpose_x_0, transpose_y = x_605_transpose_y_0, x = q_with_bias_v_cast_fp16, y = op_5414_to_fp16_quantized)[name = string("x_605_cast_fp16")];
+            tensor<int32, [8]> x_607_pad_0 = const()[name = string("x_607_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_607_mode_0 = const()[name = string("x_607_mode_0"), val = string("constant")];
+            fp16 const_378_to_fp16 = const()[name = string("const_378_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 56, 196]> x_607_cast_fp16 = pad(constant_val = const_378_to_fp16, mode = x_607_mode_0, pad = x_607_pad_0, x = x_605_cast_fp16)[name = string("x_607_cast_fp16")];
+            tensor<int32, [4]> var_5422 = const()[name = string("op_5422"), val = tensor<int32, [4]>([1, 8, -1, 56])];
+            tensor<fp16, [1, 8, 196, 56]> x_609_cast_fp16 = reshape(shape = var_5422, x = x_607_cast_fp16)[name = string("x_609_cast_fp16")];
+            tensor<int32, [4]> var_5426_begin_0 = const()[name = string("op_5426_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5426_end_0 = const()[name = string("op_5426_end_0"), val = tensor<int32, [4]>([1, 8, 196, 56])];
+            tensor<bool, [4]> var_5426_end_mask_0 = const()[name = string("op_5426_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 195, 56]> var_5426_cast_fp16 = slice_by_index(begin = var_5426_begin_0, end = var_5426_end_0, end_mask = var_5426_end_mask_0, x = x_609_cast_fp16)[name = string("op_5426_cast_fp16")];
+            tensor<int32, [4]> var_5427 = const()[name = string("op_5427"), val = tensor<int32, [4]>([1, 8, 56, 195])];
+            tensor<fp16, [1, 8, 56, 195]> matrix_bd_93_cast_fp16 = reshape(shape = var_5427, x = var_5426_cast_fp16)[name = string("matrix_bd_93_cast_fp16")];
+            bool matrix_ac_transpose_x_0 = const()[name = string("matrix_ac_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_transpose_y_0 = const()[name = string("matrix_ac_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_142_perm_0 = const()[name = string("transpose_142_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_143_perm_0 = const()[name = string("transpose_143_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 98]> transpose_143 = transpose(perm = transpose_143_perm_0, x = k_93_cast_fp16)[name = string("transpose_153")];
+            tensor<fp16, [1, 8, 56, 128]> transpose_142 = transpose(perm = transpose_142_perm_0, x = var_5410_cast_fp16)[name = string("transpose_154")];
+            tensor<fp16, [1, 8, 56, 98]> matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_142, y = transpose_143)[name = string("matrix_ac_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_begin_0 = const()[name = string("matrix_bd_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_end_0 = const()[name = string("matrix_bd_end_0"), val = tensor<int32, [4]>([1, 8, 56, 98])];
+            tensor<bool, [4]> matrix_bd_end_mask_0 = const()[name = string("matrix_bd_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 56, 98]> matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_93_cast_fp16)[name = string("matrix_bd_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_5436_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = string("op_5436_cast_fp16")];
+            fp16 _inversed_scores_93_y_0_to_fp16 = const()[name = string("_inversed_scores_93_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 56, 98]> _inversed_scores_93_cast_fp16 = mul(x = var_5436_cast_fp16, y = _inversed_scores_93_y_0_to_fp16)[name = string("_inversed_scores_93_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> scores_cast_fp16 = select(a = var_44_to_fp16, b = _inversed_scores_93_cast_fp16, cond = mask_11)[name = string("scores_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> var_5442_cast_fp16 = softmax(axis = var_58, x = scores_cast_fp16)[name = string("op_5442_cast_fp16")];
+            tensor<fp16, [1, 8, 56, 98]> input_1237_cast_fp16 = select(a = var_43_to_fp16, b = var_5442_cast_fp16, cond = mask_11)[name = string("input_1237_cast_fp16")];
+            bool x_611_transpose_x_0 = const()[name = string("x_611_transpose_x_0"), val = bool(false)];
+            bool x_611_transpose_y_0 = const()[name = string("x_611_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 98, 128]> value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = string("transpose_152")];
+            tensor<fp16, [1, 8, 56, 128]> x_611_cast_fp16 = matmul(transpose_x = x_611_transpose_x_0, transpose_y = x_611_transpose_y_0, x = input_1237_cast_fp16, y = value_cast_fp16)[name = string("x_611_cast_fp16")];
+            tensor<int32, [4]> var_5446_perm_0 = const()[name = string("op_5446_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5447 = const()[name = string("op_5447"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 56, 8, 128]> var_5446_cast_fp16 = transpose(perm = var_5446_perm_0, x = x_611_cast_fp16)[name = string("transpose_151")];
+            tensor<fp16, [1, 56, 1024]> input_1239_cast_fp16 = reshape(shape = var_5447, x = var_5446_cast_fp16)[name = string("input_1239_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535692864)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537790080)))];
+            tensor<fp16, [1, 56, 1024]> linear_214_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_out_weight_to_fp16, x = input_1239_cast_fp16)[name = string("linear_214_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1243_cast_fp16 = add(x = input_1233_cast_fp16, y = linear_214_cast_fp16)[name = string("input_1243_cast_fp16")];
+            tensor<int32, [1]> x_615_axes_0 = const()[name = string("x_615_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537792192)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537794304)))];
+            tensor<fp16, [1, 56, 1024]> x_615_cast_fp16 = layer_norm(axes = x_615_axes_0, beta = encoder_layers_23_norm_conv_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_23_norm_conv_weight_to_fp16, x = input_1243_cast_fp16)[name = string("x_615_cast_fp16")];
+            tensor<int32, [3]> input_1245_perm_0 = const()[name = string("input_1245_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1247_pad_type_0 = const()[name = string("input_1247_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1247_strides_0 = const()[name = string("input_1247_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1247_pad_0 = const()[name = string("input_1247_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1247_dilations_0 = const()[name = string("input_1247_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1247_groups_0 = const()[name = string("input_1247_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537796416))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539893632))))[name = string("encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> input_1245_cast_fp16 = transpose(perm = input_1245_perm_0, x = x_615_cast_fp16)[name = string("transpose_150")];
+            tensor<fp16, [1, 2048, 56]> input_1247_cast_fp16 = conv(dilations = input_1247_dilations_0, groups = input_1247_groups_0, pad = input_1247_pad_0, pad_type = input_1247_pad_type_0, strides = input_1247_strides_0, weight = encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1245_cast_fp16)[name = string("input_1247_cast_fp16")];
+            int32 x_617_split_num_splits_0 = const()[name = string("x_617_split_num_splits_0"), val = int32(2)];
+            int32 x_617_split_axis_0 = const()[name = string("x_617_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 56]> x_617_split_cast_fp16_0, tensor<fp16, [1, 1024, 56]> x_617_split_cast_fp16_1 = split(axis = x_617_split_axis_0, num_splits = x_617_split_num_splits_0, x = input_1247_cast_fp16)[name = string("x_617_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_617_split_1_sigmoid_cast_fp16 = sigmoid(x = x_617_split_cast_fp16_1)[name = string("x_617_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> x_617_cast_fp16 = mul(x = x_617_split_cast_fp16_0, y = x_617_split_1_sigmoid_cast_fp16)[name = string("x_617_cast_fp16")];
+            tensor<fp16, [1, 1024, 56]> input_1249_cast_fp16 = select(a = var_43_to_fp16, b = x_617_cast_fp16, cond = var_574)[name = string("input_1249_cast_fp16")];
+            bool new_x_interleave_0 = const()[name = string("new_x_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 64]> new_x_cast_fp16 = concat(axis = var_58, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_1249_cast_fp16))[name = string("new_x_cast_fp16")];
+            tensor<int32, [3]> cache_last_time_cur_begin_0 = const()[name = string("cache_last_time_cur_begin_0"), val = tensor<int32, [3]>([0, 0, 56])];
+            tensor<int32, [3]> cache_last_time_cur_end_0 = const()[name = string("cache_last_time_cur_end_0"), val = tensor<int32, [3]>([1, 1024, 64])];
+            tensor<bool, [3]> cache_last_time_cur_end_mask_0 = const()[name = string("cache_last_time_cur_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> cache_last_time_cur_cast_fp16 = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = new_x_cast_fp16)[name = string("cache_last_time_cur_cast_fp16")];
+            string x_619_pad_type_0 = const()[name = string("x_619_pad_type_0"), val = string("valid")];
+            int32 x_619_groups_0 = const()[name = string("x_619_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_619_strides_0 = const()[name = string("x_619_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_619_pad_0 = const()[name = string("x_619_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_619_dilations_0 = const()[name = string("x_619_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539897792))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539907072))))[name = string("encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_619_cast_fp16 = conv(dilations = x_619_dilations_0, groups = x_619_groups_0, pad = x_619_pad_0, pad_type = x_619_pad_type_0, strides = x_619_strides_0, weight = encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_cast_fp16)[name = string("x_619_cast_fp16")];
+            tensor<int32, [3]> input_1251_perm_0 = const()[name = string("input_1251_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_621_axes_0 = const()[name = string("x_621_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539909184)))];
+            tensor<fp16, [1024]> encoder_layers_23_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539911296)))];
+            tensor<fp16, [1, 56, 1024]> input_1251_cast_fp16 = transpose(perm = input_1251_perm_0, x = x_619_cast_fp16)[name = string("transpose_149")];
+            tensor<fp16, [1, 56, 1024]> x_621_cast_fp16 = layer_norm(axes = x_621_axes_0, beta = encoder_layers_23_conv_batch_norm_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_23_conv_batch_norm_weight_to_fp16, x = input_1251_cast_fp16)[name = string("x_621_cast_fp16")];
+            tensor<int32, [3]> input_1253_perm_0 = const()[name = string("input_1253_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 56]> input_1253_cast_fp16 = transpose(perm = input_1253_perm_0, x = x_621_cast_fp16)[name = string("transpose_148")];
+            tensor<fp16, [1, 1024, 56]> input_1255_cast_fp16 = silu(x = input_1253_cast_fp16)[name = string("input_1255_cast_fp16")];
+            string x_623_pad_type_0 = const()[name = string("x_623_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_623_strides_0 = const()[name = string("x_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_623_pad_0 = const()[name = string("x_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_623_dilations_0 = const()[name = string("x_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_623_groups_0 = const()[name = string("x_623_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(539913408))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540962048))))[name = string("encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 56]> x_623_cast_fp16 = conv(dilations = x_623_dilations_0, groups = x_623_groups_0, pad = x_623_pad_0, pad_type = x_623_pad_type_0, strides = x_623_strides_0, weight = encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1255_cast_fp16)[name = string("x_623_cast_fp16")];
+            tensor<int32, [3]> input_1257_perm_0 = const()[name = string("input_1257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 56, 1024]> input_1257_cast_fp16 = transpose(perm = input_1257_perm_0, x = x_623_cast_fp16)[name = string("transpose_147")];
+            tensor<fp16, [1, 56, 1024]> input_1259_cast_fp16 = add(x = input_1243_cast_fp16, y = input_1257_cast_fp16)[name = string("input_1259_cast_fp16")];
+            tensor<int32, [1]> input_1261_axes_0 = const()[name = string("input_1261_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540964160)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540966272)))];
+            tensor<fp16, [1, 56, 1024]> input_1261_cast_fp16 = layer_norm(axes = input_1261_axes_0, beta = encoder_layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_23_norm_feed_forward2_weight_to_fp16, x = input_1259_cast_fp16)[name = string("input_1261_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_23_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540968384)))];
+            tensor<fp16, [4096]> encoder_layers_23_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549357056)))];
+            tensor<fp16, [1, 56, 4096]> linear_215_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear1_weight_to_fp16, x = input_1261_cast_fp16)[name = string("linear_215_cast_fp16")];
+            tensor<fp16, [1, 56, 4096]> input_1265_cast_fp16 = silu(x = linear_215_cast_fp16)[name = string("input_1265_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_23_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549365312)))];
+            tensor<fp16, [1024]> encoder_layers_23_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557753984)))];
+            tensor<fp16, [1, 56, 1024]> linear_216_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear2_weight_to_fp16, x = input_1265_cast_fp16)[name = string("linear_216_cast_fp16")];
+            fp16 var_5529_to_fp16 = const()[name = string("op_5529_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 56, 1024]> var_5530_cast_fp16 = mul(x = linear_216_cast_fp16, y = var_5529_to_fp16)[name = string("op_5530_cast_fp16")];
+            tensor<fp16, [1, 56, 1024]> input_1271_cast_fp16 = add(x = input_1259_cast_fp16, y = var_5530_cast_fp16)[name = string("input_1271_cast_fp16")];
+            tensor<int32, [1]> audio_signal_axes_0 = const()[name = string("audio_signal_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557756096)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557758208)))];
+            tensor<fp16, [1, 56, 1024]> audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_23_norm_out_bias_to_fp16, epsilon = var_41_to_fp16, gamma = encoder_layers_23_norm_out_weight_to_fp16, x = input_1271_cast_fp16)[name = string("audio_signal_cast_fp16")];
+            int32 obj_5_axis_0 = const()[name = string("obj_5_axis_0"), val = int32(0)];
+            tensor<fp16, [24, 1, 56, 1024]> obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_483_cast_fp16, var_696_cast_fp16, var_909_cast_fp16, var_1122_cast_fp16, var_1335_cast_fp16, var_1548_cast_fp16, var_1761_cast_fp16, var_1974_cast_fp16, var_2187_cast_fp16, var_2400_cast_fp16, var_2613_cast_fp16, var_2826_cast_fp16, var_3039_cast_fp16, var_3252_cast_fp16, var_3465_cast_fp16, var_3678_cast_fp16, var_3891_cast_fp16, var_4104_cast_fp16, var_4317_cast_fp16, var_4530_cast_fp16, var_4743_cast_fp16, var_4956_cast_fp16, var_5169_cast_fp16, cache_last_channel_cur_cast_fp16))[name = string("obj_5_cast_fp16")];
+            int32 obj_7_axis_0 = const()[name = string("obj_7_axis_0"), val = int32(0)];
+            tensor<fp16, [24, 1, 1024, 8]> obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_587_cast_fp16, var_800_cast_fp16, var_1013_cast_fp16, var_1226_cast_fp16, var_1439_cast_fp16, var_1652_cast_fp16, var_1865_cast_fp16, var_2078_cast_fp16, var_2291_cast_fp16, var_2504_cast_fp16, var_2717_cast_fp16, var_2930_cast_fp16, var_3143_cast_fp16, var_3356_cast_fp16, var_3569_cast_fp16, var_3782_cast_fp16, var_3995_cast_fp16, var_4208_cast_fp16, var_4421_cast_fp16, var_4634_cast_fp16, var_4847_cast_fp16, var_5060_cast_fp16, var_5273_cast_fp16, cache_last_time_cur_cast_fp16))[name = string("obj_7_cast_fp16")];
+            tensor<int32, [1]> var_5546 = add(x = cache_len, y = max_audio_length_1)[name = string("op_5546")];
+            string var_5546_promoted_to_fp16_dtype_0 = const()[name = string("op_5546_promoted_to_fp16_dtype_0"), val = string("fp16")];
+            fp16 const_384_to_fp16 = const()[name = string("const_384_to_fp16"), val = fp16(-inf)];
+            fp16 var_48_promoted_to_fp16 = const()[name = string("op_48_promoted_to_fp16"), val = fp16(0x1.5p+5)];
+            tensor<fp16, [1]> var_5546_to_fp16 = cast(dtype = var_5546_promoted_to_fp16_dtype_0, x = var_5546)[name = string("cast_6")];
+            tensor<fp16, [1]> clip_1_cast_fp16 = clip(alpha = const_384_to_fp16, beta = var_48_promoted_to_fp16, x = var_5546_to_fp16)[name = string("clip_1_cast_fp16")];
+            tensor<int32, [4]> var_5568_begin_0 = const()[name = string("op_5568_begin_0"), val = tensor<int32, [4]>([0, 0, 14, 0])];
+            tensor<int32, [4]> var_5568_end_0 = const()[name = string("op_5568_end_0"), val = tensor<int32, [4]>([24, 1, 56, 1024])];
+            tensor<bool, [4]> var_5568_end_mask_0 = const()[name = string("op_5568_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [24, 1, 42, 1024]> var_5568_cast_fp16 = slice_by_index(begin = var_5568_begin_0, end = var_5568_end_0, end_mask = var_5568_end_mask_0, x = obj_5_cast_fp16)[name = string("op_5568_cast_fp16")];
+            int32 var_5588_one_hot_vector_size_0 = const()[name = string("op_5588_one_hot_vector_size_0"), val = int32(128)];
+            int32 var_5588_axis_0 = const()[name = string("op_5588_axis_0"), val = int32(-1)];
+            int32 var_5588_on_value_0 = const()[name = string("op_5588_on_value_0"), val = int32(1)];
+            int32 var_5588_off_value_0 = const()[name = string("op_5588_off_value_0"), val = int32(0)];
+            tensor<int32, [1, 128]> var_5588 = one_hot(axis = var_5588_axis_0, indices = prompt_id, off_value = var_5588_off_value_0, on_value = var_5588_on_value_0, one_hot_vector_size = var_5588_one_hot_vector_size_0)[name = string("op_5588")];
+            tensor<int32, [1]> var_5591_axes_0 = const()[name = string("op_5591_axes_0"), val = tensor<int32, [1]>([1])];
+            string cast_245_to_fp16_dtype_0 = const()[name = string("cast_245_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 128]> var_5588_to_fp16 = cast(dtype = cast_245_to_fp16_dtype_0, x = var_5588)[name = string("cast_5")];
+            tensor<fp16, [1, 1, 128]> var_5591_cast_fp16 = expand_dims(axes = var_5591_axes_0, x = var_5588_to_fp16)[name = string("op_5591_cast_fp16")];
+            tensor<int32, [3]> one_hot_reps_0 = const()[name = string("one_hot_reps_0"), val = tensor<int32, [3]>([1, 56, 1])];
+            tensor<fp16, [1, 56, 128]> one_hot_cast_fp16 = tile(reps = one_hot_reps_0, x = var_5591_cast_fp16)[name = string("one_hot_cast_fp16")];
+            int32 var_5600 = const()[name = string("op_5600"), val = int32(-1)];
+            bool input_1273_interleave_0 = const()[name = string("input_1273_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 56, 1152]> input_1273_cast_fp16 = concat(axis = var_5600, interleave = input_1273_interleave_0, values = (audio_signal_cast_fp16, one_hot_cast_fp16))[name = string("input_1273_cast_fp16")];
+            tensor<fp16, [2048, 1152]> prompt_kernel_0_weight_to_fp16 = const()[name = string("prompt_kernel_0_weight_to_fp16"), val = tensor<fp16, [2048, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557760320)))];
+            tensor<fp16, [2048]> prompt_kernel_0_bias_to_fp16 = const()[name = string("prompt_kernel_0_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562478976)))];
+            tensor<fp16, [1, 56, 2048]> linear_217_cast_fp16 = linear(bias = prompt_kernel_0_bias_to_fp16, weight = prompt_kernel_0_weight_to_fp16, x = input_1273_cast_fp16)[name = string("linear_217_cast_fp16")];
+            tensor<fp16, [1, 56, 2048]> input_cast_fp16 = relu(x = linear_217_cast_fp16)[name = string("input_cast_fp16")];
+            tensor<fp16, [1024, 2048]> prompt_kernel_2_weight_to_fp16 = const()[name = string("prompt_kernel_2_weight_to_fp16"), val = tensor<fp16, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562483136)))];
+            tensor<fp16, [1024]> prompt_kernel_2_bias_to_fp16 = const()[name = string("prompt_kernel_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(566677504)))];
+            tensor<fp16, [1, 56, 1024]> linear_218_cast_fp16 = linear(bias = prompt_kernel_2_bias_to_fp16, weight = prompt_kernel_2_weight_to_fp16, x = input_cast_fp16)[name = string("linear_218_cast_fp16")];
+            tensor<int32, [3]> var_5613_perm_0 = const()[name = string("op_5613_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string var_5613_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5613_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            string cast_246_dtype_0 = const()[name = string("cast_246_dtype_0"), val = string("int32")];
+            tensor<int32, [4]> var_5621_perm_0 = const()[name = string("op_5621_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string var_5621_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5621_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [4]> var_5624_perm_0 = const()[name = string("op_5624_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string var_5624_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5624_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            string cast_247_dtype_0 = const()[name = string("cast_247_dtype_0"), val = string("int32")];
+            tensor<int32, [1]> cache_len_out = cast(dtype = cast_247_dtype_0, x = clip_1_cast_fp16)[name = string("cast_0")];
+            tensor<fp16, [1, 24, 1024, 8]> var_5624_cast_fp16 = transpose(perm = var_5624_perm_0, x = obj_7_cast_fp16)[name = string("transpose_144")];
+            tensor<fp32, [1, 24, 1024, 8]> cache_time_out = cast(dtype = var_5624_cast_fp16_to_fp32_dtype_0, x = var_5624_cast_fp16)[name = string("cast_1")];
+            tensor<fp16, [1, 24, 42, 1024]> var_5621_cast_fp16 = transpose(perm = var_5621_perm_0, x = var_5568_cast_fp16)[name = string("transpose_145")];
+            tensor<fp32, [1, 24, 42, 1024]> cache_channel_out = cast(dtype = var_5621_cast_fp16_to_fp32_dtype_0, x = var_5621_cast_fp16)[name = string("cast_2")];
+            tensor<int32, [1]> encoded_length = cast(dtype = cast_246_dtype_0, x = clip_0_cast_fp16)[name = string("cast_3")];
+            tensor<fp16, [1, 1024, 56]> var_5613_cast_fp16 = transpose(perm = var_5613_perm_0, x = linear_218_cast_fp16)[name = string("transpose_146")];
+            tensor<fp32, [1, 1024, 56]> encoded = cast(dtype = var_5613_cast_fp16_to_fp32_dtype_0, x = var_5613_cast_fp16)[name = string("cast_4")];
+        } -> (encoded, encoded_length, cache_channel_out, cache_time_out, cache_len_out);
+}
\ No newline at end of file
diff --git a/fr/4480ms/encoder.mlmodelc/weights/weight.bin b/fr/4480ms/encoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..e770743e8f69de769852fc42baf0304b71ade176
--- /dev/null
+++ b/fr/4480ms/encoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a88ef3a70bab53d01290a8c9d3f29c171843a29356f93213384700ab3525ab0
+size 566679616
diff --git a/fr/4480ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/4480ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..2bccf0128a76f1c624788bb613ec4ba49eaf1241
--- /dev/null
+++ b/fr/4480ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9812c206b63aefeb42a47341acc3f9dadb0c6e71f64706a282114bf0a60373db
+size 793485
diff --git a/fr/4480ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/4480ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..e770743e8f69de769852fc42baf0304b71ade176
--- /dev/null
+++ b/fr/4480ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a88ef3a70bab53d01290a8c9d3f29c171843a29356f93213384700ab3525ab0
+size 566679616
diff --git a/fr/4480ms/encoder.mlpackage/Manifest.json b/fr/4480ms/encoder.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..a8bdbffb50f301171c7ac13d8f900bc840438643
--- /dev/null
+++ b/fr/4480ms/encoder.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "0CEB2AD9-5203-4556-962D-6FB2C0D7D1E9": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "5EBD8374-043E-43F4-BCD6-D98F95B69791": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "5EBD8374-043E-43F4-BCD6-D98F95B69791"
+}
diff --git a/fr/4480ms/joint.mlmodelc/analytics/coremldata.bin b/fr/4480ms/joint.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..38aade026cc9376471863f9f09be292c9fe8578b
--- /dev/null
+++ b/fr/4480ms/joint.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad1f6f5571ee9e7ad8a533b4679ed7416879f4d8e5fc10f40835c1d014b822a4
+size 243
diff --git a/fr/4480ms/joint.mlmodelc/coremldata.bin b/fr/4480ms/joint.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..f33da57e5d96472b26d6ba7a76e961d6d469b79b
--- /dev/null
+++ b/fr/4480ms/joint.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9c243bc2047e7c9be1b75e3c8efdd9d983f782e38436753b851d8fd4515a21dd
+size 341
diff --git a/fr/4480ms/joint.mlmodelc/model.mil b/fr/4480ms/joint.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..041dd87f66df1673f6226d5c642c93a64d4e2730
--- /dev/null
+++ b/fr/4480ms/joint.mlmodelc/model.mil
@@ -0,0 +1,31 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp32, [1, 640, 1]> decoder, tensor<fp32, [1, 1024, 1]> encoder) {
+            tensor<int32, [3]> input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [640, 1024]> module_enc_weight_to_fp16 = const()[name = string("module_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [640]> module_enc_bias_to_fp16 = const()[name = string("module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1310848)))];
+            tensor<fp16, [1, 1024, 1]> encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_2")];
+            tensor<fp16, [1, 1, 1024]> input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_to_fp16)[name = string("transpose_1")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = module_enc_bias_to_fp16, weight = module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> module_pred_weight_to_fp16 = const()[name = string("module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1312192)))];
+            tensor<fp16, [640]> module_pred_bias_to_fp16 = const()[name = string("module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2131456)))];
+            tensor<fp16, [1, 640, 1]> decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_1")];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_to_fp16)[name = string("transpose_0")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = module_pred_bias_to_fp16, weight = module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_23_axes_0 = const()[name = string("op_23_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = string("op_23_cast_fp16")];
+            tensor<int32, [1]> var_25_axes_0 = const()[name = string("op_25_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_25_cast_fp16 = expand_dims(axes = var_25_axes_0, x = linear_1_cast_fp16)[name = string("op_25_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_25_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("input_7_cast_fp16")];
+            tensor<fp16, [849, 640]> module_joint_net_2_weight_to_fp16 = const()[name = string("module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2132800)))];
+            tensor<fp16, [849]> module_joint_net_2_bias_to_fp16 = const()[name = string("module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [849]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3219584)))];
+            tensor<fp16, [1, 1, 1, 849]> linear_2_cast_fp16 = linear(bias = module_joint_net_2_bias_to_fp16, weight = module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_2_cast_fp16")];
+            string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 1, 1, 849]> logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_0")];
+        } -> (logits);
+}
\ No newline at end of file
diff --git a/fr/4480ms/joint.mlmodelc/weights/weight.bin b/fr/4480ms/joint.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..109a5bf4eb27d11fcd040ea350a8e6d00d18932a
--- /dev/null
+++ b/fr/4480ms/joint.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9fa5fb0d14d87484d9ffd1f4d30996ea456c14ea3e5ab9d910486d8e86d3a540
+size 3221346
diff --git a/fr/4480ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/4480ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..0c650fe487d18d0d937a1fc295a3eee53b4aacbe
--- /dev/null
+++ b/fr/4480ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4bac39a8a21b3e7e8cdeba1d03b92e8f743bf87cb2de3b1d85c505a580a0b218
+size 4486
diff --git a/fr/4480ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/4480ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..109a5bf4eb27d11fcd040ea350a8e6d00d18932a
--- /dev/null
+++ b/fr/4480ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9fa5fb0d14d87484d9ffd1f4d30996ea456c14ea3e5ab9d910486d8e86d3a540
+size 3221346
diff --git a/fr/4480ms/joint.mlpackage/Manifest.json b/fr/4480ms/joint.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7e49a7390ba8e327aff397061f0fd4476b7c70a
--- /dev/null
+++ b/fr/4480ms/joint.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "0DC18C6B-B3DE-4D3A-80EE-28CAA06F6708": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "88105E4A-1C49-40F4-A835-C9B725CEEC59": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "88105E4A-1C49-40F4-A835-C9B725CEEC59"
+}
diff --git a/fr/4480ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin b/fr/4480ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..a856ad184cad37664eb30bb699b3a88e6401000d
--- /dev/null
+++ b/fr/4480ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28b5c13f8ce2d8154edb31968c174d322ee4ae76db42f41ecf7ac5e0fd87f57d
+size 243
diff --git a/fr/4480ms/joint_noencproj_batched.mlmodelc/coremldata.bin b/fr/4480ms/joint_noencproj_batched.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..2d98aad58fa9306bb5d9eb65992a468896e119ef
--- /dev/null
+++ b/fr/4480ms/joint_noencproj_batched.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4803b38b79692a3dbd10cb227f06d6f41308b1a00e4c795a8d53688ae71c83fd
+size 406
diff --git a/fr/4480ms/joint_noencproj_batched.mlmodelc/model.mil b/fr/4480ms/joint_noencproj_batched.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..b66479ae4eaa92652a98a34a18e993259a14e6c0
--- /dev/null
+++ b/fr/4480ms/joint_noencproj_batched.mlmodelc/model.mil
@@ -0,0 +1,26 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios18>(tensor<fp32, [1, 640, 1]> decoder, tensor<fp32, [1, 4, 640]> encoder_proj) {
+            tensor<int32, [3]> input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819328)))];
+            tensor<fp16, [1, 640, 1]> decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_2")];
+            tensor<fp16, [1, 1, 640]> input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = decoder_to_fp16)[name = string("transpose_0")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [1]> var_15_axes_0 = const()[name = string("op_15_axes_0"), val = tensor<int32, [1]>([2])];
+            string encoder_proj_to_fp16_dtype_0 = const()[name = string("encoder_proj_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 4, 640]> encoder_proj_to_fp16 = cast(dtype = encoder_proj_to_fp16_dtype_0, x = encoder_proj)[name = string("cast_1")];
+            tensor<fp16, [1, 4, 1, 640]> var_15_cast_fp16 = expand_dims(axes = var_15_axes_0, x = encoder_proj_to_fp16)[name = string("op_15_cast_fp16")];
+            tensor<int32, [1]> var_17_axes_0 = const()[name = string("op_17_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_17_cast_fp16 = expand_dims(axes = var_17_axes_0, x = linear_0_cast_fp16)[name = string("op_17_cast_fp16")];
+            tensor<fp16, [1, 4, 1, 640]> input_3_cast_fp16 = add(x = var_15_cast_fp16, y = var_17_cast_fp16)[name = string("input_3_cast_fp16")];
+            tensor<fp16, [1, 4, 1, 640]> input_5_cast_fp16 = relu(x = input_3_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<fp16, [849, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(820672)))];
+            tensor<fp16, [849]> joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [849]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1907456)))];
+            tensor<fp16, [1, 4, 1, 849]> linear_1_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_5_cast_fp16)[name = string("linear_1_cast_fp16")];
+            string linear_1_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_1_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 4, 1, 849]> logits = cast(dtype = linear_1_cast_fp16_to_fp32_dtype_0, x = linear_1_cast_fp16)[name = string("cast_0")];
+        } -> (logits);
+}
\ No newline at end of file
diff --git a/fr/4480ms/joint_noencproj_batched.mlmodelc/weights/weight.bin b/fr/4480ms/joint_noencproj_batched.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..4e66049a9a8bde2b92859df3192b671ddf9ab81d
--- /dev/null
+++ b/fr/4480ms/joint_noencproj_batched.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9a9dd24e6b63b97540b0db0e990418bc6fddc9c2bf85b82eb226e6df1ea02a4
+size 1909218
diff --git a/fr/4480ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/4480ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..4f82db74749dc3350042c3f6c1453c4d4e620d50
--- /dev/null
+++ b/fr/4480ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d13f08c4961187e47001773d7d4d5d5937ba7ff68475eb7c1bac41d7014c64a1
+size 3839
diff --git a/fr/4480ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/4480ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..4e66049a9a8bde2b92859df3192b671ddf9ab81d
--- /dev/null
+++ b/fr/4480ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9a9dd24e6b63b97540b0db0e990418bc6fddc9c2bf85b82eb226e6df1ea02a4
+size 1909218
diff --git a/fr/4480ms/joint_noencproj_batched.mlpackage/Manifest.json b/fr/4480ms/joint_noencproj_batched.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..5f2e3080d9f6dc5617b9630128af453e31562411
--- /dev/null
+++ b/fr/4480ms/joint_noencproj_batched.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "EFDDA9F1-A088-4CE2-8A64-5928EB8F429C": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "F28FF7B3-AC0A-4ACA-93F8-DF201DFFE75B": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "F28FF7B3-AC0A-4ACA-93F8-DF201DFFE75B"
+}
diff --git a/fr/4480ms/metadata.json b/fr/4480ms/metadata.json
new file mode 100644
index 0000000000000000000000000000000000000000..0e5db9343e6483f68ebefda519f445405125f79b
--- /dev/null
+++ b/fr/4480ms/metadata.json
@@ -0,0 +1,198 @@
+{
+  "model": "nvidia/nemotron-asr-streaming-multilingual-0.6b",
+  "model_class": "nemo.collections.asr.models.rnnt_bpe_models_prompt.EncDecRNNTBPEModelWithPrompt",
+  "sample_rate": 16000,
+  "mel_features": 128,
+  "chunk_mel_frames": 448,
+  "pre_encode_cache": 9,
+  "total_mel_frames": 457,
+  "att_context_size": [
+    42,
+    13
+  ],
+  "vocab_size": 848,
+  "blank_idx": 848,
+  "vocab_pruned": true,
+  "vocab_pruned_original_size": 13087,
+  "cache_channel_shape": [
+    1,
+    24,
+    42,
+    1024
+  ],
+  "cache_time_shape": [
+    1,
+    24,
+    1024,
+    8
+  ],
+  "decoder_hidden": 640,
+  "decoder_layers": 2,
+  "encoder_dim": 1024,
+  "num_prompts": 128,
+  "prompt_dictionary": {
+    "af-ZA": 54,
+    "am-ET": 49,
+    "ar": 7,
+    "ar-AR": 7,
+    "auto": 101,
+    "ay-BO": 81,
+    "az-AZ": 66,
+    "bg": 30,
+    "bg-BG": 30,
+    "bn-IN": 36,
+    "cs": 22,
+    "cs-CZ": 22,
+    "da": 25,
+    "da-DK": 25,
+    "de": 9,
+    "de-DE": 9,
+    "el": 21,
+    "el-GR": 21,
+    "en": 0,
+    "en-GB": 1,
+    "en-US": 0,
+    "enGB": 1,
+    "es": 3,
+    "es-ES": 2,
+    "es-US": 3,
+    "esES": 2,
+    "et": 60,
+    "et-EE": 60,
+    "fa-IR": 38,
+    "fi": 26,
+    "fi-FI": 26,
+    "fr": 8,
+    "fr-CA": 100,
+    "fr-FR": 8,
+    "gn-PY": 82,
+    "gu-IN": 42,
+    "ha-NG": 50,
+    "haw-US": 97,
+    "he-IL": 64,
+    "hi": 6,
+    "hi-HI": 6,
+    "hi-IN": 6,
+    "hr": 29,
+    "hr-HR": 29,
+    "hu": 23,
+    "hu-HU": 23,
+    "hy-AM": 68,
+    "id-ID": 34,
+    "ig-NG": 53,
+    "it": 15,
+    "it-IT": 15,
+    "ja-JA": 10,
+    "ja-JP": 10,
+    "ka-GE": 67,
+    "km-KH": 47,
+    "kn-IN": 43,
+    "ko": 14,
+    "ko-KO": 14,
+    "ko-KR": 14,
+    "ku-TR": 65,
+    "ky-KG": 71,
+    "ln-CD": 58,
+    "lt": 31,
+    "lt-LT": 31,
+    "lv": 61,
+    "lv-LV": 61,
+    "mi-NZ": 96,
+    "ml-IN": 44,
+    "mr-IN": 41,
+    "ms-MY": 35,
+    "mt-MT": 102,
+    "nah-MX": 83,
+    "nb": 103,
+    "nb-NO": 103,
+    "ne-NP": 46,
+    "nl": 16,
+    "nl-NL": 16,
+    "nn": 104,
+    "nn-NO": 104,
+    "no": 27,
+    "no-NO": 27,
+    "ny-MW": 57,
+    "or-KE": 59,
+    "pl": 17,
+    "pl-PL": 17,
+    "pt": 13,
+    "pt-BR": 12,
+    "pt-PT": 13,
+    "qu-PE": 80,
+    "ro": 20,
+    "ro-RO": 20,
+    "ru": 11,
+    "ru-RU": 11,
+    "rw-RW": 55,
+    "si-LK": 45,
+    "sk": 28,
+    "sk-SK": 28,
+    "sl": 62,
+    "sl-SI": 62,
+    "sm-WS": 98,
+    "so-SO": 56,
+    "sv": 24,
+    "sv-SE": 24,
+    "sw-KE": 48,
+    "ta-IN": 39,
+    "te-IN": 40,
+    "tg-TJ": 70,
+    "th-TH": 32,
+    "to-TO": 99,
+    "tr": 18,
+    "tr-TR": 18,
+    "uk": 19,
+    "uk-UA": 19,
+    "ur-PK": 37,
+    "uz-UZ": 69,
+    "vi-VN": 33,
+    "yo-NG": 52,
+    "zh-CN": 4,
+    "zh-TW": 5,
+    "zh-ZH": 4,
+    "zu-ZA": 51
+  },
+  "default_prompt_id": 101,
+  "lang_tag_token_ids": [
+    1,
+    72,
+    122,
+    170,
+    230,
+    231,
+    263,
+    282,
+    390,
+    415,
+    430,
+    483,
+    501,
+    519,
+    542,
+    565,
+    594,
+    619,
+    620,
+    629,
+    640,
+    659,
+    660,
+    677,
+    704,
+    717,
+    735,
+    741,
+    790,
+    791,
+    793,
+    794,
+    795,
+    800,
+    819,
+    829,
+    830,
+    844,
+    845
+  ]
+}
\ No newline at end of file
diff --git a/fr/4480ms/preprocessor.mlmodelc/analytics/coremldata.bin b/fr/4480ms/preprocessor.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..e80c07518c7261d08beb9bc52cd276894a4bd14c
--- /dev/null
+++ b/fr/4480ms/preprocessor.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:12264cf4116cb366a0a4cc606ffba954ad2f9494d25776e76112c294b847ffb7
+size 243
diff --git a/fr/4480ms/preprocessor.mlmodelc/coremldata.bin b/fr/4480ms/preprocessor.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..0944be2be6b80940760d1f5f5f0f11ac817288bb
--- /dev/null
+++ b/fr/4480ms/preprocessor.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7b11e08aba46d1845d8ad3f247717e0f6fae35b21d71d52e44a69ea73587bfe
+size 371
diff --git a/fr/4480ms/preprocessor.mlmodelc/model.mil b/fr/4480ms/preprocessor.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..0b8261362f9cbf465b530a0d2d0ee9a2b2f462cd
--- /dev/null
+++ b/fr/4480ms/preprocessor.mlmodelc/model.mil
@@ -0,0 +1,122 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp32, [1, ?]> audio, tensor<int32, [1]> audio_length) [FlexibleShapeInformation = tuple<tuple<string, dict<string, tensor<int32, [?]>>>, tuple<string, dict<string, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"audio", [1, 1]}}), ("RangeDims", {{"audio", [[1, 1], [1, 480000]]}})))] {
+            int32 var_9 = const()[name = string("op_9"), val = int32(1)];
+            int32 var_10 = const()[name = string("op_10"), val = int32(160)];
+            int32 var_12 = const()[name = string("op_12"), val = int32(0)];
+            int32 var_33 = const()[name = string("op_33"), val = int32(512)];
+            tensor<int32, [1]> var_34 = add(x = audio_length, y = var_33)[name = string("op_34")];
+            int32 var_35 = const()[name = string("op_35"), val = int32(512)];
+            tensor<int32, [1]> var_36 = sub(x = var_34, y = var_35)[name = string("op_36")];
+            tensor<int32, [1]> floor_div_0 = floor_div(x = var_36, y = var_10)[name = string("floor_div_0")];
+            tensor<bool, [1]> var_39 = equal(x = audio_length, y = var_12)[name = string("op_39")];
+            tensor<int32, [1]> var_40 = const()[name = string("op_40"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> mel_length = select(a = var_40, b = floor_div_0, cond = var_39)[name = string("seq_len")];
+            string audio_to_fp16_dtype_0 = const()[name = string("audio_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, ?]> audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = string("cast_14")];
+            tensor<int32, [2]> var_42_shape_cast_fp16 = shape(x = audio_to_fp16)[name = string("op_42_shape_cast_fp16")];
+            int32 gather_0_axis_0 = const()[name = string("gather_0_axis_0"), val = int32(0)];
+            int32 gather_0_batch_dims_0 = const()[name = string("gather_0_batch_dims_0"), val = int32(0)];
+            bool gather_0_validate_indices_0 = const()[name = string("gather_0_validate_indices_0"), val = bool(false)];
+            string var_42_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_42_shape_cast_fp16_to_int16_dtype_0"), val = string("int16")];
+            uint16 select_0_to_uint16 = const()[name = string("select_0_to_uint16"), val = uint16(1)];
+            tensor<int16, [2]> var_42_shape_cast_fp16_to_int16 = cast(dtype = var_42_shape_cast_fp16_to_int16_dtype_0, x = var_42_shape_cast_fp16)[name = string("cast_13")];
+            int16 gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_42_shape_cast_fp16_to_int16)[name = string("gather_0_cast_uint16")];
+            string gather_0_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_0_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            int32 const_0 = const()[name = string("const_0"), val = int32(0)];
+            int32 const_1 = const()[name = string("const_1"), val = int32(1)];
+            int32 gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = string("cast_12")];
+            tensor<int32, [?]> var_43 = range_1d(end = gather_0_cast_uint16_to_int32, start = const_0, step = const_1)[name = string("op_43")];
+            tensor<int32, [1]> var_44_axes_0 = const()[name = string("op_44_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> var_44 = expand_dims(axes = var_44_axes_0, x = var_43)[name = string("op_44")];
+            tensor<int32, [1]> var_45_axes_0 = const()[name = string("op_45_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_45 = expand_dims(axes = var_45_axes_0, x = audio_length)[name = string("op_45")];
+            tensor<bool, [1, ?]> timemask = less(x = var_44, y = var_45)[name = string("timemask")];
+            tensor<int32, [2]> var_48_begin_0 = const()[name = string("op_48_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_48_end_0 = const()[name = string("op_48_end_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<bool, [2]> var_48_end_mask_0 = const()[name = string("op_48_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<bool, [2]> var_48_squeeze_mask_0 = const()[name = string("op_48_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
+            tensor<fp16, [1]> var_48_cast_fp16 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, squeeze_mask = var_48_squeeze_mask_0, x = audio_to_fp16)[name = string("op_48_cast_fp16")];
+            tensor<int32, [1]> var_49_axes_0 = const()[name = string("op_49_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1]> var_49_cast_fp16 = expand_dims(axes = var_49_axes_0, x = var_48_cast_fp16)[name = string("op_49_cast_fp16")];
+            tensor<int32, [2]> var_51_begin_0 = const()[name = string("op_51_begin_0"), val = tensor<int32, [2]>([0, 1])];
+            tensor<int32, [2]> var_51_end_0 = const()[name = string("op_51_end_0"), val = tensor<int32, [2]>([1, 0])];
+            tensor<bool, [2]> var_51_end_mask_0 = const()[name = string("op_51_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<fp16, [1, ?]> var_51_cast_fp16 = slice_by_index(begin = var_51_begin_0, end = var_51_end_0, end_mask = var_51_end_mask_0, x = audio_to_fp16)[name = string("op_51_cast_fp16")];
+            tensor<int32, [2]> var_53_begin_0 = const()[name = string("op_53_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_53_end_0 = const()[name = string("op_53_end_0"), val = tensor<int32, [2]>([1, -1])];
+            tensor<bool, [2]> var_53_end_mask_0 = const()[name = string("op_53_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp16, [1, ?]> var_53_cast_fp16 = slice_by_index(begin = var_53_begin_0, end = var_53_end_0, end_mask = var_53_end_mask_0, x = audio_to_fp16)[name = string("op_53_cast_fp16")];
+            fp16 var_54_to_fp16 = const()[name = string("op_54_to_fp16"), val = fp16(0x1.f0cp-1)];
+            tensor<fp16, [1, ?]> var_55_cast_fp16 = mul(x = var_53_cast_fp16, y = var_54_to_fp16)[name = string("op_55_cast_fp16")];
+            tensor<fp16, [1, ?]> var_56_cast_fp16 = sub(x = var_51_cast_fp16, y = var_55_cast_fp16)[name = string("op_56_cast_fp16")];
+            bool x_3_interleave_0 = const()[name = string("x_3_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, ?]> x_3_cast_fp16 = concat(axis = var_9, interleave = x_3_interleave_0, values = (var_49_cast_fp16, var_56_cast_fp16))[name = string("x_3_cast_fp16")];
+            tensor<bool, [1, ?]> var_59 = logical_not(x = timemask)[name = string("op_59")];
+            fp16 var_16_to_fp16 = const()[name = string("op_16_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, ?]> input_1_cast_fp16 = select(a = var_16_to_fp16, b = x_3_cast_fp16, cond = var_59)[name = string("input_1_cast_fp16")];
+            tensor<int32, [3]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, ?]> input_3_cast_fp16 = reshape(shape = concat_1x, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")];
+            tensor<int32, [6]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 256, 256])];
+            string input_5_mode_0 = const()[name = string("input_5_mode_0"), val = string("constant")];
+            fp16 const_3_to_fp16 = const()[name = string("const_3_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, ?]> input_5_cast_fp16 = pad(constant_val = const_3_to_fp16, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<int32, [2]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [2]>([1, -1])];
+            tensor<fp16, [1, ?]> input_cast_fp16 = reshape(shape = concat_2x, x = input_5_cast_fp16)[name = string("input_cast_fp16")];
+            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([160])];
+            tensor<int32, [1]> expand_dims_4_axes_0 = const()[name = string("expand_dims_4_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, ?]> expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = input_cast_fp16)[name = string("expand_dims_4_cast_fp16")];
+            string conv_0_pad_type_0 = const()[name = string("conv_0_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> conv_0_pad_0 = const()[name = string("conv_0_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_0_dilations_0 = const()[name = string("conv_0_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 conv_0_groups_0 = const()[name = string("conv_0_groups_0"), val = int32(1)];
+            tensor<fp16, [257, 1, 512]> expand_dims_1_to_fp16 = const()[name = string("expand_dims_1_to_fp16"), val = tensor<fp16, [257, 1, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 257, ?]> conv_0_cast_fp16 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_0_cast_fp16")];
+            string conv_1_pad_type_0 = const()[name = string("conv_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> conv_1_pad_0 = const()[name = string("conv_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_1_dilations_0 = const()[name = string("conv_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 conv_1_groups_0 = const()[name = string("conv_1_groups_0"), val = int32(1)];
+            tensor<fp16, [257, 1, 512]> expand_dims_2_to_fp16 = const()[name = string("expand_dims_2_to_fp16"), val = tensor<fp16, [257, 1, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263296)))];
+            tensor<fp16, [1, 257, ?]> conv_1_cast_fp16 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_1_cast_fp16")];
+            int32 stack_0_axis_0 = const()[name = string("stack_0_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 257, ?, 2]> stack_0_cast_fp16 = stack(axis = stack_0_axis_0, values = (conv_0_cast_fp16, conv_1_cast_fp16))[name = string("stack_0_cast_fp16")];
+            fp16 var_19_promoted_to_fp16 = const()[name = string("op_19_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1, 257, ?, 2]> var_74_cast_fp16 = pow(x = stack_0_cast_fp16, y = var_19_promoted_to_fp16)[name = string("op_74_cast_fp16")];
+            tensor<int32, [1]> var_76_axes_0 = const()[name = string("op_76_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_76_keep_dims_0 = const()[name = string("op_76_keep_dims_0"), val = bool(false)];
+            tensor<fp16, [1, 257, ?]> var_76_cast_fp16 = reduce_sum(axes = var_76_axes_0, keep_dims = var_76_keep_dims_0, x = var_74_cast_fp16)[name = string("op_76_cast_fp16")];
+            tensor<fp16, [1, 257, ?]> x_11_cast_fp16 = identity(x = var_76_cast_fp16)[name = string("x_11_cast_fp16")];
+            bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)];
+            bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 128, 257]> const_4_to_fp16 = const()[name = string("const_4_to_fp16"), val = tensor<fp16, [1, 128, 257]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526528)))];
+            tensor<fp16, [1, 128, ?]> x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = const_4_to_fp16, y = x_11_cast_fp16)[name = string("x_13_cast_fp16")];
+            fp16 var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 128, ?]> var_84_cast_fp16 = add(x = x_13_cast_fp16, y = var_83_to_fp16)[name = string("op_84_cast_fp16")];
+            fp32 x_epsilon_0 = const()[name = string("x_epsilon_0"), val = fp32(0x1p-149)];
+            tensor<fp16, [1, 128, ?]> x_cast_fp16 = log(epsilon = x_epsilon_0, x = var_84_cast_fp16)[name = string("x_cast_fp16")];
+            tensor<int32, [3]> var_86_shape_cast_fp16 = shape(x = x_cast_fp16)[name = string("op_86_shape_cast_fp16")];
+            int32 gather_5_axis_0 = const()[name = string("gather_5_axis_0"), val = int32(0)];
+            int32 gather_5_batch_dims_0 = const()[name = string("gather_5_batch_dims_0"), val = int32(0)];
+            bool gather_5_validate_indices_0 = const()[name = string("gather_5_validate_indices_0"), val = bool(false)];
+            string var_86_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_86_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
+            uint16 select_5_to_uint16 = const()[name = string("select_5_to_uint16"), val = uint16(2)];
+            tensor<uint16, [3]> var_86_shape_cast_fp16_to_uint16 = cast(dtype = var_86_shape_cast_fp16_to_uint16_dtype_0, x = var_86_shape_cast_fp16)[name = string("cast_11")];
+            uint16 gather_5_cast_uint16 = gather(axis = gather_5_axis_0, batch_dims = gather_5_batch_dims_0, indices = select_5_to_uint16, validate_indices = gather_5_validate_indices_0, x = var_86_shape_cast_fp16_to_uint16)[name = string("gather_5_cast_uint16")];
+            string gather_5_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_5_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            int32 const_5 = const()[name = string("const_5"), val = int32(0)];
+            int32 const_6 = const()[name = string("const_6"), val = int32(1)];
+            int32 gather_5_cast_uint16_to_int32 = cast(dtype = gather_5_cast_uint16_to_int32_dtype_0, x = gather_5_cast_uint16)[name = string("cast_10")];
+            tensor<int32, [?]> mask_1 = range_1d(end = gather_5_cast_uint16_to_int32, start = const_5, step = const_6)[name = string("mask_1")];
+            tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = string("expand_dims_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = string("expand_dims_0")];
+            tensor<int32, [1]> var_91_axes_0 = const()[name = string("op_91_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_91 = expand_dims(axes = var_91_axes_0, x = mel_length)[name = string("op_91")];
+            tensor<bool, [1, ?]> mask = greater_equal(x = expand_dims_0, y = var_91)[name = string("mask")];
+            tensor<int32, [1]> var_93_axes_0 = const()[name = string("op_93_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, ?]> var_93 = expand_dims(axes = var_93_axes_0, x = mask)[name = string("op_93")];
+            tensor<fp16, [1, 128, ?]> processed_signal_cast_fp16 = select(a = var_16_to_fp16, b = x_cast_fp16, cond = var_93)[name = string("processed_signal_cast_fp16")];
+            string processed_signal_cast_fp16_to_fp32_dtype_0 = const()[name = string("processed_signal_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 128, ?]> mel = cast(dtype = processed_signal_cast_fp16_to_fp32_dtype_0, x = processed_signal_cast_fp16)[name = string("cast_9")];
+        } -> (mel, mel_length);
+}
\ No newline at end of file
diff --git a/fr/4480ms/preprocessor.mlmodelc/weights/weight.bin b/fr/4480ms/preprocessor.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411
--- /dev/null
+++ b/fr/4480ms/preprocessor.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3
+size 592384
diff --git a/fr/4480ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/4480ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..d1f9d01bb6fca81fc2a3212d0fd405343a5c5971
--- /dev/null
+++ b/fr/4480ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:daad3e5dc1a13a1dc66997f9c1c1280b68646d7ccd7b6359fa8c361d63b5c09b
+size 15878
diff --git a/fr/4480ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/4480ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411
--- /dev/null
+++ b/fr/4480ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3
+size 592384
diff --git a/fr/4480ms/preprocessor.mlpackage/Manifest.json b/fr/4480ms/preprocessor.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..9722979403b2b2f0a821e990b4205cca52694033
--- /dev/null
+++ b/fr/4480ms/preprocessor.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "11EFF503-1D75-4BC3-9480-44E97E36229D": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "A4EAB1E5-DDC2-4445-9D00-C29BCDE18F72": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "A4EAB1E5-DDC2-4445-9D00-C29BCDE18F72"
+}
diff --git a/fr/4480ms/tokenizer.json b/fr/4480ms/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..53c1afe82c2a72808c78fffb846345d590c453fd
--- /dev/null
+++ b/fr/4480ms/tokenizer.json
@@ -0,0 +1,851 @@
+{
+  "0": "<unk>",
+  "1": "<bg-BG>",
+  "2": "▁",
+  "3": ".",
+  "4": ",",
+  "5": "e",
+  "6": "t",
+  "7": "a",
+  "8": "s",
+  "9": "o",
+  "10": "i",
+  "11": "r",
+  "12": "l",
+  "13": "u",
+  "14": "d",
+  "15": "c",
+  "16": "h",
+  "17": "m",
+  "18": "p",
+  "19": "n",
+  "20": "g",
+  "21": "f",
+  "22": "en",
+  "23": "in",
+  "24": "on",
+  "25": "y",
+  "26": "er",
+  "27": "an",
+  "28": "w",
+  "29": "0",
+  "30": "b",
+  "31": "v",
+  "32": "2",
+  "33": "1",
+  "34": "▁the",
+  "35": "k",
+  "36": "3",
+  "37": "5",
+  "38": "q",
+  "39": "I",
+  "40": "é",
+  "41": "4",
+  "42": "z",
+  "43": "A",
+  "44": "6",
+  "45": "j",
+  "46": "E",
+  "47": "7",
+  "48": "T",
+  "49": "8",
+  "50": "9",
+  "51": "S",
+  "52": "x",
+  "53": "C",
+  "54": "M",
+  "55": "P",
+  "56": "D",
+  "57": "B",
+  "58": "U",
+  "59": "N",
+  "60": "O",
+  "61": "G",
+  "62": "à",
+  "63": "L",
+  "64": "è",
+  "65": "V",
+  "66": "R",
+  "67": "ó",
+  "68": "J",
+  "69": "K",
+  "70": "ê",
+  "71": "ç",
+  "72": "<cs-CZ>",
+  "73": "st",
+  "74": "ch",
+  "75": "le",
+  "76": "li",
+  "77": "▁po",
+  "78": "no",
+  "79": "to",
+  "80": "me",
+  "81": "te",
+  "82": "ho",
+  "83": "▁pro",
+  "84": "ro",
+  "85": "▁na",
+  "86": "ce",
+  "87": "la",
+  "88": "ni",
+  "89": "ra",
+  "90": "ti",
+  "91": "lo",
+  "92": "ko",
+  "93": "po",
+  "94": "je",
+  "95": "de",
+  "96": "na",
+  "97": "mi",
+  "98": "ci",
+  "99": "▁by",
+  "100": "ve",
+  "101": "▁za",
+  "102": "▁A",
+  "103": "re",
+  "104": "ou",
+  "105": "vo",
+  "106": "né",
+  "107": "va",
+  "108": "mo",
+  "109": "ze",
+  "110": "ne",
+  "111": "ka",
+  "112": "▁jak",
+  "113": "ny",
+  "114": "vi",
+  "115": "ent",
+  "116": "▁pan",
+  "117": "Z",
+  "118": "▁To",
+  "119": "▁Je",
+  "120": "ă",
+  "121": "X",
+  "122": "<da-DK>",
+  "123": "▁for",
+  "124": "▁at",
+  "125": "et",
+  "126": "▁vi",
+  "127": "al",
+  "128": "▁de",
+  "129": "▁der",
+  "130": "or",
+  "131": "om",
+  "132": "and",
+  "133": "at",
+  "134": "▁af",
+  "135": "ge",
+  "136": "ar",
+  "137": "is",
+  "138": "▁be",
+  "139": "un",
+  "140": "lig",
+  "141": "▁man",
+  "142": "ig",
+  "143": "▁som",
+  "144": "el",
+  "145": "ag",
+  "146": "erne",
+  "147": "▁den",
+  "148": "ste",
+  "149": "id",
+  "150": "ske",
+  "151": "iv",
+  "152": "▁her",
+  "153": "ion",
+  "154": "am",
+  "155": "ur",
+  "156": "for",
+  "157": "▁pr",
+  "158": "▁sig",
+  "159": "▁ind",
+  "160": "ende",
+  "161": "▁Vi",
+  "162": "ation",
+  "163": "mme",
+  "164": "tage",
+  "165": "▁op",
+  "166": "▁fra",
+  "167": "▁alle",
+  "168": "▁Men",
+  "169": "▁var",
+  "170": "<de-DE>",
+  "171": "▁die",
+  "172": "sch",
+  "173": "it",
+  "174": "gen",
+  "175": "▁W",
+  "176": "▁B",
+  "177": "▁E",
+  "178": "▁F",
+  "179": "ll",
+  "180": "▁es",
+  "181": "▁K",
+  "182": "ie",
+  "183": "au",
+  "184": "ich",
+  "185": "ck",
+  "186": "ten",
+  "187": "mal",
+  "188": "ein",
+  "189": "▁T",
+  "190": "▁mit",
+  "191": "hr",
+  "192": "ter",
+  "193": "tz",
+  "194": "▁G",
+  "195": "ben",
+  "196": "um",
+  "197": "us",
+  "198": "cht",
+  "199": "il",
+  "200": "ut",
+  "201": "▁ver",
+  "202": "▁hier",
+  "203": "▁von",
+  "204": "ri",
+  "205": "ach",
+  "206": "ol",
+  "207": "▁Da",
+  "208": "sp",
+  "209": "ell",
+  "210": "▁ja",
+  "211": "uch",
+  "212": "wi",
+  "213": "rei",
+  "214": "▁Ge",
+  "215": "und",
+  "216": "isch",
+  "217": "▁sie",
+  "218": "▁Ja",
+  "219": "▁du",
+  "220": "▁dem",
+  "221": "▁Sch",
+  "222": "▁Ma",
+  "223": "▁De",
+  "224": "▁Le",
+  "225": "▁In",
+  "226": "▁Re",
+  "227": "▁sein",
+  "228": "▁Mi",
+  "229": "▁Ha",
+  "230": "<el-GR>",
+  "231": "<et-EE>",
+  "232": "ma",
+  "233": "ta",
+  "234": "se",
+  "235": "da",
+  "236": "si",
+  "237": "ks",
+  "238": "ga",
+  "239": "he",
+  "240": "mu",
+  "241": "tu",
+  "242": "ha",
+  "243": "ja",
+  "244": "gi",
+  "245": "nd",
+  "246": "oo",
+  "247": "gu",
+  "248": "ju",
+  "249": "est",
+  "250": "▁pa",
+  "251": "tud",
+  "252": "▁pea",
+  "253": "sta",
+  "254": "di",
+  "255": "mise",
+  "256": "▁su",
+  "257": "ide",
+  "258": "val",
+  "259": "▁Me",
+  "260": "ment",
+  "261": "▁meil",
+  "262": "Q",
+  "263": "<fi-FI>",
+  "264": "ssa",
+  "265": "lla",
+  "266": "ty",
+  "267": "ki",
+  "268": "pa",
+  "269": "lle",
+  "270": "lu",
+  "271": "tta",
+  "272": "isi",
+  "273": "ise",
+  "274": "tte",
+  "275": "ista",
+  "276": "inen",
+  "277": "llis",
+  "278": "vu",
+  "279": "▁voi",
+  "280": "iden",
+  "281": "ë",
+  "282": "<fr-FR>",
+  "283": "▁est",
+  "284": "▁c",
+  "285": "▁que",
+  "286": "▁à",
+  "287": "es",
+  "288": "▁un",
+  "289": "▁pas",
+  "290": "▁les",
+  "291": "▁qui",
+  "292": "▁il",
+  "293": "▁vous",
+  "294": "▁des",
+  "295": "▁qu",
+  "296": "▁pour",
+  "297": "▁par",
+  "298": "▁ça",
+  "299": "▁une",
+  "300": "ant",
+  "301": "ais",
+  "302": "ez",
+  "303": "▁dans",
+  "304": "▁C",
+  "305": "tre",
+  "306": "ir",
+  "307": "elle",
+  "308": "eur",
+  "309": "▁sur",
+  "310": "▁con",
+  "311": "▁Et",
+  "312": "ement",
+  "313": "tion",
+  "314": "té",
+  "315": "▁tout",
+  "316": "mp",
+  "317": "ique",
+  "318": "▁plus",
+  "319": "eux",
+  "320": "▁dé",
+  "321": "▁fait",
+  "322": "qu",
+  "323": "▁comme",
+  "324": "ac",
+  "325": "ait",
+  "326": "che",
+  "327": "▁mais",
+  "328": "que",
+  "329": "ul",
+  "330": "▁avec",
+  "331": "▁bien",
+  "332": "age",
+  "333": "▁mon",
+  "334": "end",
+  "335": "▁faire",
+  "336": "▁être",
+  "337": "ver",
+  "338": "▁peu",
+  "339": "▁même",
+  "340": "tra",
+  "341": "cha",
+  "342": "▁nous",
+  "343": "▁donc",
+  "344": "▁sont",
+  "345": "▁moi",
+  "346": "ille",
+  "347": "ff",
+  "348": "▁ex",
+  "349": "▁Il",
+  "350": "▁très",
+  "351": "▁cette",
+  "352": "im",
+  "353": "ité",
+  "354": "▁dire",
+  "355": "▁peut",
+  "356": "ance",
+  "357": "aire",
+  "358": "mé",
+  "359": "▁app",
+  "360": "▁aussi",
+  "361": "▁petit",
+  "362": "aux",
+  "363": "▁parce",
+  "364": "onne",
+  "365": "mb",
+  "366": "man",
+  "367": "▁On",
+  "368": "▁quand",
+  "369": "▁autre",
+  "370": "ô",
+  "371": "▁chose",
+  "372": "▁puis",
+  "373": "▁était",
+  "374": "ndre",
+  "375": "port",
+  "376": "▁vraiment",
+  "377": "ence",
+  "378": "▁Mais",
+  "379": "î",
+  "380": "▁avoir",
+  "381": "form",
+  "382": "▁faut",
+  "383": "ign",
+  "384": "▁où",
+  "385": "près",
+  "386": "▁beaucoup",
+  "387": "ture",
+  "388": "û",
+  "389": "â",
+  "390": "<hu-HU>",
+  "391": "▁az",
+  "392": "ok",
+  "393": "gy",
+  "394": "ek",
+  "395": "és",
+  "396": "em",
+  "397": "▁is",
+  "398": "os",
+  "399": "ban",
+  "400": "ik",
+  "401": "cs",
+  "402": "él",
+  "403": "ér",
+  "404": "nek",
+  "405": "ség",
+  "406": "én",
+  "407": "hat",
+  "408": "▁ha",
+  "409": "ép",
+  "410": "vé",
+  "411": "leg",
+  "412": "▁ami",
+  "413": "É",
+  "414": "rend",
+  "415": "<hr-HR>",
+  "416": "▁bi",
+  "417": "▁sa",
+  "418": "ru",
+  "419": "go",
+  "420": "sti",
+  "421": "▁pri",
+  "422": "ima",
+  "423": "nu",
+  "424": "▁pre",
+  "425": "nja",
+  "426": "zi",
+  "427": "vr",
+  "428": "ca",
+  "429": "ba",
+  "430": "<it-IT>",
+  "431": "▁di",
+  "432": "▁che",
+  "433": "co",
+  "434": "▁per",
+  "435": "▁non",
+  "436": "do",
+  "437": "gli",
+  "438": "so",
+  "439": "amo",
+  "440": "sa",
+  "441": "ndo",
+  "442": "fi",
+  "443": "pi",
+  "444": "nti",
+  "445": "tto",
+  "446": "tro",
+  "447": "▁fa",
+  "448": "chi",
+  "449": "▁qua",
+  "450": "bi",
+  "451": "▁del",
+  "452": "pe",
+  "453": "ssi",
+  "454": "▁sono",
+  "455": "nte",
+  "456": "sso",
+  "457": "nta",
+  "458": "cu",
+  "459": "ggi",
+  "460": "sto",
+  "461": "gra",
+  "462": "spe",
+  "463": "cco",
+  "464": "nde",
+  "465": "fe",
+  "466": "pu",
+  "467": "zza",
+  "468": "sci",
+  "469": "▁poi",
+  "470": "sco",
+  "471": "stra",
+  "472": "▁quel",
+  "473": "qui",
+  "474": "sse",
+  "475": "▁inter",
+  "476": "sce",
+  "477": "▁com",
+  "478": "▁comp",
+  "479": "▁prima",
+  "480": "▁imp",
+  "481": "▁cui",
+  "482": "scri",
+  "483": "<lt-LT>",
+  "484": "ai",
+  "485": "▁ir",
+  "486": "as",
+  "487": "▁tai",
+  "488": "uo",
+  "489": "tin",
+  "490": "▁vis",
+  "491": "ly",
+  "492": "gal",
+  "493": "jo",
+  "494": "tar",
+  "495": "yb",
+  "496": "▁Ir",
+  "497": "▁Tai",
+  "498": "▁nu",
+  "499": "▁mes",
+  "500": "▁pra",
+  "501": "<lv-LV>",
+  "502": "▁pie",
+  "503": "ies",
+  "504": "ot",
+  "505": "▁Un",
+  "506": "▁Ta",
+  "507": "▁lie",
+  "508": "dar",
+  "509": "sim",
+  "510": "dien",
+  "511": "gan",
+  "512": "▁ap",
+  "513": "▁nav",
+  "514": "▁cit",
+  "515": "▁Ne",
+  "516": "zin",
+  "517": "reiz",
+  "518": "▁20",
+  "519": "<nl-NL>",
+  "520": "▁dat",
+  "521": "▁we",
+  "522": "▁En",
+  "523": "▁dan",
+  "524": "▁met",
+  "525": "▁wat",
+  "526": "der",
+  "527": "ui",
+  "528": "den",
+  "529": "op",
+  "530": "oor",
+  "531": "▁of",
+  "532": "ven",
+  "533": "acht",
+  "534": "ond",
+  "535": "▁dit",
+  "536": "▁wil",
+  "537": "▁We",
+  "538": "vol",
+  "539": "llen",
+  "540": "▁Dan",
+  "541": "▁Nou",
+  "542": "<pl-PL>",
+  "543": "nie",
+  "544": "wa",
+  "545": "my",
+  "546": "cie",
+  "547": "nia",
+  "548": "wo",
+  "549": "rze",
+  "550": "by",
+  "551": "za",
+  "552": "dy",
+  "553": "ry",
+  "554": "mie",
+  "555": "pie",
+  "556": "uje",
+  "557": "▁bo",
+  "558": "bie",
+  "559": "sy",
+  "560": "▁Po",
+  "561": "ski",
+  "562": "▁No",
+  "563": "▁Na",
+  "564": "▁musi",
+  "565": "<pt-BR>",
+  "566": "pt",
+  "567": "-",
+  "568": "▁eu",
+  "569": "▁para",
+  "570": "▁pe",
+  "571": "▁tem",
+  "572": "▁O",
+  "573": "pre",
+  "574": "ria",
+  "575": "▁fo",
+  "576": "mos",
+  "577": "bo",
+  "578": "▁por",
+  "579": "▁Eu",
+  "580": "be",
+  "581": "▁esse",
+  "582": "ente",
+  "583": "▁mas",
+  "584": "qua",
+  "585": "fica",
+  "586": "▁vou",
+  "587": "▁Se",
+  "588": "▁fala",
+  "589": "...",
+  "590": "ço",
+  "591": "▁Por",
+  "592": "▁Co",
+  "593": "iza",
+  "594": "<ro-RO>",
+  "595": "▁mai",
+  "596": "sc",
+  "597": "are",
+  "598": "▁din",
+  "599": "rea",
+  "600": "du",
+  "601": "▁M",
+  "602": "▁fac",
+  "603": "lor",
+  "604": "▁mult",
+  "605": "per",
+  "606": "cur",
+  "607": "tor",
+  "608": "inte",
+  "609": "▁sau",
+  "610": "tat",
+  "611": "ori",
+  "612": "▁prim",
+  "613": "▁lui",
+  "614": "▁sub",
+  "615": "▁prin",
+  "616": "▁alt",
+  "617": "stru",
+  "618": "fer",
+  "619": "<ru-RU>",
+  "620": "<sk-SK>",
+  "621": "sk",
+  "622": "▁aj",
+  "623": "ob",
+  "624": "ali",
+  "625": "nov",
+  "626": "rov",
+  "627": "rob",
+  "628": "▁spo",
+  "629": "<sl-SL>",
+  "630": "sl",
+  "631": "udi",
+  "632": "del",
+  "633": "▁sem",
+  "634": "▁samo",
+  "635": "▁let",
+  "636": "▁Pre",
+  "637": "▁prot",
+  "638": "▁internet",
+  "639": "▁film",
+  "640": "<sv-SE>",
+  "641": "▁att",
+  "642": "▁inte",
+  "643": "▁av",
+  "644": "all",
+  "645": "era",
+  "646": "pp",
+  "647": "sam",
+  "648": "▁vill",
+  "649": "erna",
+  "650": "ande",
+  "651": "ade",
+  "652": "▁hur",
+  "653": "bil",
+  "654": "▁min",
+  "655": "lev",
+  "656": "land",
+  "657": "text",
+  "658": "person",
+  "659": "<uk-UA>",
+  "660": "<ar-AR>",
+  "661": "'",
+  "662": "▁H",
+  "663": "▁D",
+  "664": "aus",
+  "665": "▁N",
+  "666": "▁Be",
+  "667": "mm",
+  "668": "ab",
+  "669": "▁Er",
+  "670": "ssen",
+  "671": "hl",
+  "672": "rie",
+  "673": "▁An",
+  "674": "rau",
+  "675": "▁So",
+  "676": "▁Aus",
+  "677": "<en-GB>",
+  "678": "▁can",
+  "679": "ed",
+  "680": "ay",
+  "681": "th",
+  "682": "ic",
+  "683": "hi",
+  "684": "▁not",
+  "685": "ight",
+  "686": "ex",
+  "687": "▁four",
+  "688": "▁great",
+  "689": "▁six",
+  "690": "ill",
+  "691": "▁but",
+  "692": "▁got",
+  "693": "▁don",
+  "694": "▁out",
+  "695": "▁from",
+  "696": "▁fine",
+  "697": "able",
+  "698": "ever",
+  "699": "▁month",
+  "700": "▁question",
+  "701": "▁possible",
+  "702": "▁moment",
+  "703": "▁change",
+  "704": "<en-US>",
+  "705": "ow",
+  "706": "ive",
+  "707": "ate",
+  "708": "ad",
+  "709": "ect",
+  "710": "ng",
+  "711": "ther",
+  "712": "act",
+  "713": "ist",
+  "714": "▁part",
+  "715": "side",
+  "716": "\"",
+  "717": "<es-ES>",
+  "718": "ción",
+  "719": "res",
+  "720": "▁La",
+  "721": "dos",
+  "722": "▁las",
+  "723": "men",
+  "724": "par",
+  "725": "rio",
+  "726": "cho",
+  "727": "ble",
+  "728": "▁Ca",
+  "729": "▁Su",
+  "730": "▁son",
+  "731": "▁Con",
+  "732": "ones",
+  "733": "ecto",
+  "734": "▁Com",
+  "735": "<es-US>",
+  "736": "cia",
+  "737": "▁Y",
+  "738": "ron",
+  "739": "les",
+  "740": "bu",
+  "741": "<fr-CA>",
+  "742": "ré",
+  "743": "ée",
+  "744": "▁Les",
+  "745": "our",
+  "746": "▁Ce",
+  "747": "com",
+  "748": "▁Elle",
+  "749": "▁Cet",
+  "750": "ux",
+  "751": "ale",
+  "752": "ction",
+  "753": "▁pré",
+  "754": "▁deux",
+  "755": "if",
+  "756": "lé",
+  "757": "ère",
+  "758": "ière",
+  "759": "iste",
+  "760": "▁parti",
+  "761": "▁été",
+  "762": "avec",
+  "763": "jour",
+  "764": "app",
+  "765": "cul",
+  "766": "▁égale",
+  "767": "aine",
+  "768": "gue",
+  "769": "▁nombre",
+  "770": "▁étai",
+  "771": "tout",
+  "772": "▁grand",
+  "773": "▁commun",
+  "774": "Une",
+  "775": "œ",
+  "776": "ï",
+  "777": "À",
+  "778": "Â",
+  "779": "av",
+  "780": "ouv",
+  "781": "êt",
+  "782": "ois",
+  "783": "pri",
+  "784": "voir",
+  "785": "sion",
+  "786": "ix",
+  "787": "était",
+  "788": "ard",
+  "789": "aient",
+  "790": "<he-IL>",
+  "791": "<hi-IN>",
+  "792": "!",
+  "793": "<ja-JP>",
+  "794": "<ko-KR>",
+  "795": "<nb-NO>",
+  "796": "▁ble",
+  "797": "opp",
+  "798": "▁Han",
+  "799": "▁han",
+  "800": "<nn-NO>",
+  "801": "eg",
+  "802": "kk",
+  "803": "dig",
+  "804": "ord",
+  "805": "▁tru",
+  "806": "▁sei",
+  "807": "ller",
+  "808": "▁fle",
+  "809": "car",
+  "810": "▁foi",
+  "811": "ito",
+  "812": "ram",
+  "813": "fa",
+  "814": "▁meu",
+  "815": "nça",
+  "816": "▁mil",
+  "817": "▁passa",
+  "818": "▁casa",
+  "819": "<pt-PT>",
+  "820": "das",
+  "821": "▁Pa",
+  "822": "forma",
+  "823": "mar",
+  "824": "▁entre",
+  "825": "fun",
+  "826": "▁grande",
+  "827": "▁Este",
+  "828": "var",
+  "829": "<th-TH>",
+  "830": "<tr-TR>",
+  "831": "ş",
+  "832": "ğ",
+  "833": "ya",
+  "834": "▁ve",
+  "835": "lar",
+  "836": "ler",
+  "837": "ye",
+  "838": "▁bu",
+  "839": "lan",
+  "840": "ara",
+  "841": "▁Bu",
+  "842": "inde",
+  "843": "yo",
+  "844": "<zh-CN>",
+  "845": "<vi-VN>",
+  "846": "▁t",
+  "847": "nh",
+  "848": "<blank>"
+}
\ No newline at end of file
diff --git a/fr/560ms/decoder.mlmodelc/analytics/coremldata.bin b/fr/560ms/decoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..65b800e186dcb664ec3d46edb11fa43215dce1f7
--- /dev/null
+++ b/fr/560ms/decoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d3becf0abd7a7efd30577f33f4144cfcb8c08155fdd62903eb1a45d7b02b87c
+size 243
diff --git a/fr/560ms/decoder.mlmodelc/coremldata.bin b/fr/560ms/decoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..a921778bfd0fef1f29bd148ddcb6c74c6116663e
--- /dev/null
+++ b/fr/560ms/decoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d6b7630abf0c1b447913bf9d1ed5935d625bcb70e978c607820b2ff35ee8f07f
+size 433
diff --git a/fr/560ms/decoder.mlmodelc/model.mil b/fr/560ms/decoder.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..0de41956c1a2cac40f0629b26d6b60b23d5fe6b3
--- /dev/null
+++ b/fr/560ms/decoder.mlmodelc/model.mil
@@ -0,0 +1,64 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            int32 y_axis_0 = const()[name = string("y_axis_0"), val = int32(0)];
+            int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)];
+            bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [849, 640]> module_prediction_embed_weight_to_fp16 = const()[name = string("module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_8")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)];
+            int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)];
+            string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")];
+            int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)];
+            int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)];
+            string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            string input_lstm_layer_0_direction_0 = const()[name = string("input_lstm_layer_0_direction_0"), val = string("forward")];
+            bool input_lstm_layer_0_output_sequence_0 = const()[name = string("input_lstm_layer_0_output_sequence_0"), val = bool(true)];
+            string input_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")];
+            string input_lstm_layer_0_cell_activation_0 = const()[name = string("input_lstm_layer_0_cell_activation_0"), val = string("tanh")];
+            string input_lstm_layer_0_activation_0 = const()[name = string("input_lstm_layer_0_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1086848)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4363712)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7640576)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = string("transpose_2")];
+            tensor<fp16, [1, 1, 640]> input_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_lstm_layer_0_cast_fp16_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_lstm_layer_0_cell_activation_0, direction = input_lstm_layer_0_direction_0, initial_c = input_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_lstm_h0_squeeze_axes_0 = const()[name = string("input_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_c0_squeeze_axes_0 = const()[name = string("input_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_lstm_c0_squeeze_cast_fp16")];
+            string input_direction_0 = const()[name = string("input_direction_0"), val = string("forward")];
+            bool input_output_sequence_0 = const()[name = string("input_output_sequence_0"), val = bool(true)];
+            string input_recurrent_activation_0 = const()[name = string("input_recurrent_activation_0"), val = string("sigmoid")];
+            string input_cell_activation_0 = const()[name = string("input_cell_activation_0"), val = string("tanh")];
+            string input_activation_0 = const()[name = string("input_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7645760)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10922624)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14199488)))];
+            tensor<fp16, [1, 1, 640]> input_cast_fp16_0, tensor<fp16, [1, 640]> input_cast_fp16_1, tensor<fp16, [1, 640]> input_cast_fp16_2 = lstm(activation = input_activation_0, bias = concat_3_to_fp16, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze_cast_fp16, initial_h = input_lstm_h0_squeeze_cast_fp16, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_lstm_layer_0_cast_fp16_0)[name = string("input_cast_fp16")];
+            int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_cast_fp16_1, input_cast_fp16_1))[name = string("obj_3_cast_fp16")];
+            string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_lstm_layer_0_cast_fp16_2, input_cast_fp16_2))[name = string("obj_cast_fp16")];
+            string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
+            string transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = string("transpose_0_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp16, [1, 640, 1]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = string("transpose_1")];
+            tensor<fp32, [1, 640, 1]> decoder_out = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = string("cast_3")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")];
+        } -> (decoder_out, h_out, c_out);
+}
\ No newline at end of file
diff --git a/fr/560ms/decoder.mlmodelc/weights/weight.bin b/fr/560ms/decoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..82ea1696ddad052267574c52c3e3b0bccde479f5
--- /dev/null
+++ b/fr/560ms/decoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ef2aeef85ed664eb60850330a014c5c299325ba1047d690654dc7ffa870e95a
+size 14204672
diff --git a/fr/560ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/560ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..c09a4248bce6ae30d0023f4ec349ae462e19aad7
--- /dev/null
+++ b/fr/560ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c82c0fbdef3af45d82f9b1fe84c0c822154020a5b3cd0e5070d33e9ef0beeca6
+size 10359
diff --git a/fr/560ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/560ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..82ea1696ddad052267574c52c3e3b0bccde479f5
--- /dev/null
+++ b/fr/560ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ef2aeef85ed664eb60850330a014c5c299325ba1047d690654dc7ffa870e95a
+size 14204672
diff --git a/fr/560ms/decoder.mlpackage/Manifest.json b/fr/560ms/decoder.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..80bff9abaf12e957f68dbde257eaf82079fed11b
--- /dev/null
+++ b/fr/560ms/decoder.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "08617906-EB26-428B-BCF5-BC6013603E2C": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "89EF9907-E7AF-4633-B6D7-C8D472B636BC": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "89EF9907-E7AF-4633-B6D7-C8D472B636BC"
+}
diff --git a/fr/560ms/decoder_joint.mlmodelc/analytics/coremldata.bin b/fr/560ms/decoder_joint.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..22debc561a233f97bd7fde99dc9d1a31fb0231de
--- /dev/null
+++ b/fr/560ms/decoder_joint.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f8367de6fb06c7b8053f5ecbbc8e4c7e5100f85f1e1c9ba6e9d9b3e31ecca3d
+size 243
diff --git a/fr/560ms/decoder_joint.mlmodelc/coremldata.bin b/fr/560ms/decoder_joint.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..54e005ff86e2274927ae7ef36a7e0ee4bb128c3b
--- /dev/null
+++ b/fr/560ms/decoder_joint.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0566656fc89cd81234bb22159c3ccebfb8c0277e3b740d12a8fb53288c1efcd
+size 454
diff --git a/fr/560ms/decoder_joint.mlmodelc/model.mil b/fr/560ms/decoder_joint.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..1438ac2ef0ec1f9fc664a6752938e369f7de8e9d
--- /dev/null
+++ b/fr/560ms/decoder_joint.mlmodelc/model.mil
@@ -0,0 +1,83 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [1, 1024, 1]> encoder, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            int32 y_axis_0 = const()[name = string("y_axis_0"), val = int32(0)];
+            int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)];
+            bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [849, 640]> decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_9")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16 = gather(axis = y_axis_0, batch_dims = y_batch_dims_0, indices = token_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)];
+            int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)];
+            string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_8")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")];
+            int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)];
+            int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)];
+            string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")];
+            bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)];
+            string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")];
+            string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1086848)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4363712)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7640576)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16)[name = string("transpose_4")];
+            tensor<fp16, [1, 1, 640]> input_5_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")];
+            string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")];
+            bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)];
+            string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")];
+            string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7645760)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10922624)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14199488)))];
+            tensor<fp16, [1, 1, 640]> input_5_cast_fp16_0, tensor<fp16, [1, 640]> input_5_cast_fp16_1, tensor<fp16, [1, 640]> input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")];
+            int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")];
+            string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")];
+            string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [3]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, [3]> input_7_perm_0 = const()[name = string("input_7_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [640, 1024]> joint_module_enc_weight_to_fp16 = const()[name = string("joint_module_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14204672)))];
+            tensor<fp16, [640]> joint_module_enc_bias_to_fp16 = const()[name = string("joint_module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15515456)))];
+            tensor<fp16, [1, 1024, 1]> encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_4")];
+            tensor<fp16, [1, 1, 1024]> input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = encoder_to_fp16)[name = string("transpose_2")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_enc_bias_to_fp16, weight = joint_module_enc_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15516800)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16336064)))];
+            tensor<fp16, [1, 1, 640]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_3")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_79_axes_0 = const()[name = string("op_79_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_79_cast_fp16 = expand_dims(axes = var_79_axes_0, x = linear_0_cast_fp16)[name = string("op_79_cast_fp16")];
+            tensor<int32, [1]> var_80_axes_0 = const()[name = string("op_80_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_80_cast_fp16 = expand_dims(axes = var_80_axes_0, x = linear_1_cast_fp16)[name = string("op_80_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_11_cast_fp16 = add(x = var_79_cast_fp16, y = var_80_cast_fp16)[name = string("input_11_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_13_cast_fp16 = relu(x = input_11_cast_fp16)[name = string("input_13_cast_fp16")];
+            tensor<fp16, [849, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16337408)))];
+            tensor<fp16, [849]> joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [849]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17424192)))];
+            tensor<fp16, [1, 1, 1, 849]> linear_2_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_2_cast_fp16")];
+            string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 1, 1, 849]> logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_3")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_5")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_6")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")];
+        } -> (logits, h_out, c_out);
+}
\ No newline at end of file
diff --git a/fr/560ms/decoder_joint.mlmodelc/weights/weight.bin b/fr/560ms/decoder_joint.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..7e387cabbe194b0f606c14bdf2b03b19d697a8f3
--- /dev/null
+++ b/fr/560ms/decoder_joint.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39ce7eecdfb6a432f391c931412d9a5cd15fb0ef85562ef0c207e4a18fe6b50e
+size 17425954
diff --git a/fr/560ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/560ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..8de2f672a8b40625e6b0f13654d4cba788880b07
--- /dev/null
+++ b/fr/560ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72ce6705a73b8b0488f31dd35a385908473aff8bef0ffe290f56bb51068a1d26
+size 13745
diff --git a/fr/560ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/560ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..7e387cabbe194b0f606c14bdf2b03b19d697a8f3
--- /dev/null
+++ b/fr/560ms/decoder_joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39ce7eecdfb6a432f391c931412d9a5cd15fb0ef85562ef0c207e4a18fe6b50e
+size 17425954
diff --git a/fr/560ms/decoder_joint.mlpackage/Manifest.json b/fr/560ms/decoder_joint.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..967aa622f9a5ac2f95978e9feda32d61c8a73643
--- /dev/null
+++ b/fr/560ms/decoder_joint.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "8FCEBC80-AF4D-42E1-B70D-96BB47913BC7": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "A77788EC-CA36-4DF3-BB1B-E6182E0ABF28": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "8FCEBC80-AF4D-42E1-B70D-96BB47913BC7"
+}
diff --git a/fr/560ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin b/fr/560ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..425b9fbc36cf7602d11ebba8a2e29be972229238
--- /dev/null
+++ b/fr/560ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:878825967d0c03e7b016222a7b824badb35d1d7e134ad3a973abc00dc5fe353a
+size 243
diff --git a/fr/560ms/decoder_joint_noencproj.mlmodelc/coremldata.bin b/fr/560ms/decoder_joint_noencproj.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..479d8bc5a3eab6af7208785446fe73a569a877dc
--- /dev/null
+++ b/fr/560ms/decoder_joint_noencproj.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ebb7868572e2691055cf8d926fe61e2049bdc0a3e6349e82227f0b82dd849814
+size 519
diff --git a/fr/560ms/decoder_joint_noencproj.mlmodelc/model.mil b/fr/560ms/decoder_joint_noencproj.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..ef9fa8f205f75bf414b61d0b622d2976b441fe5b
--- /dev/null
+++ b/fr/560ms/decoder_joint_noencproj.mlmodelc/model.mil
@@ -0,0 +1,91 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios18>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [1, 1, 640]> encoder_proj, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            int32 y_batch_dims_0 = const()[name = string("y_batch_dims_0"), val = int32(0)];
+            bool y_validate_indices_0 = const()[name = string("y_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [849, 640]> decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string token_to_int16_dtype_0 = const()[name = string("token_to_int16_dtype_0"), val = string("int16")];
+            string cast_1_dtype_0 = const()[name = string("cast_1_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
+            tensor<int16, [1, 1]> token_to_int16 = cast(dtype = token_to_int16_dtype_0, x = token)[name = string("cast_10")];
+            tensor<int32, [1, 1]> cast_1 = cast(dtype = cast_1_dtype_0, x = token_to_int16)[name = string("cast_9")];
+            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
+            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(849)];
+            tensor<int32, [1, 1]> add_2 = add(x = cast_1, y = slice_by_index_0)[name = string("add_2")];
+            tensor<int32, [1, 1]> select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = string("select_0")];
+            int32 y_cast_fp16_cast_uint16_axis_0 = const()[name = string("y_cast_fp16_cast_uint16_axis_0"), val = int32(0)];
+            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
+            tensor<int16, [1, 1]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_8")];
+            tensor<fp16, [1, 1, 640]> y_cast_fp16_cast_uint16_cast_uint16 = gather(axis = y_cast_fp16_cast_uint16_axis_0, batch_dims = y_batch_dims_0, indices = select_0_to_int16, validate_indices = y_validate_indices_0, x = decoder_module_prediction_embed_weight_to_fp16)[name = string("y_cast_fp16_cast_uint16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            int32 split_0_num_splits_0 = const()[name = string("split_0_num_splits_0"), val = int32(2)];
+            int32 split_0_axis_0 = const()[name = string("split_0_axis_0"), val = int32(0)];
+            string h_in_to_fp16_dtype_0 = const()[name = string("h_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = string("cast_7")];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in_to_fp16)[name = string("split_0_cast_fp16")];
+            int32 split_1_num_splits_0 = const()[name = string("split_1_num_splits_0"), val = int32(2)];
+            int32 split_1_axis_0 = const()[name = string("split_1_axis_0"), val = int32(0)];
+            string c_in_to_fp16_dtype_0 = const()[name = string("c_in_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = string("cast_6")];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in_to_fp16)[name = string("split_1_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = string("input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            string input_5_lstm_layer_0_direction_0 = const()[name = string("input_5_lstm_layer_0_direction_0"), val = string("forward")];
+            bool input_5_lstm_layer_0_output_sequence_0 = const()[name = string("input_5_lstm_layer_0_output_sequence_0"), val = bool(true)];
+            string input_5_lstm_layer_0_recurrent_activation_0 = const()[name = string("input_5_lstm_layer_0_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_lstm_layer_0_cell_activation_0 = const()[name = string("input_5_lstm_layer_0_cell_activation_0"), val = string("tanh")];
+            string input_5_lstm_layer_0_activation_0 = const()[name = string("input_5_lstm_layer_0_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = string("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1086848)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = string("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4363712)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = string("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7640576)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = string("transpose_3")];
+            tensor<fp16, [1, 1, 640]> input_5_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> input_5_lstm_layer_0_cast_fp16_2 = lstm(activation = input_5_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = input_5_lstm_layer_0_cell_activation_0, direction = input_5_lstm_layer_0_direction_0, initial_c = input_5_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_lstm_layer_0_output_sequence_0, recurrent_activation = input_5_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = string("input_5_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_h0_squeeze_axes_0 = const()[name = string("input_5_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_h0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = string("input_5_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_5_lstm_c0_squeeze_axes_0 = const()[name = string("input_5_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> input_5_lstm_c0_squeeze_cast_fp16 = squeeze(axes = input_5_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = string("input_5_lstm_c0_squeeze_cast_fp16")];
+            string input_5_direction_0 = const()[name = string("input_5_direction_0"), val = string("forward")];
+            bool input_5_output_sequence_0 = const()[name = string("input_5_output_sequence_0"), val = bool(true)];
+            string input_5_recurrent_activation_0 = const()[name = string("input_5_recurrent_activation_0"), val = string("sigmoid")];
+            string input_5_cell_activation_0 = const()[name = string("input_5_cell_activation_0"), val = string("tanh")];
+            string input_5_activation_0 = const()[name = string("input_5_activation_0"), val = string("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = string("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7645760)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = string("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10922624)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = string("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14199488)))];
+            tensor<fp16, [1, 1, 640]> input_5_cast_fp16_0, tensor<fp16, [1, 640]> input_5_cast_fp16_1, tensor<fp16, [1, 640]> input_5_cast_fp16_2 = lstm(activation = input_5_activation_0, bias = concat_3_to_fp16, cell_activation = input_5_cell_activation_0, direction = input_5_direction_0, initial_c = input_5_lstm_c0_squeeze_cast_fp16, initial_h = input_5_lstm_h0_squeeze_cast_fp16, output_sequence = input_5_output_sequence_0, recurrent_activation = input_5_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = input_5_lstm_layer_0_cast_fp16_0)[name = string("input_5_cast_fp16")];
+            int32 obj_3_axis_0 = const()[name = string("obj_3_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_3_cast_fp16 = stack(axis = obj_3_axis_0, values = (input_5_lstm_layer_0_cast_fp16_1, input_5_cast_fp16_1))[name = string("obj_3_cast_fp16")];
+            string obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_3_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> obj_cast_fp16 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")];
+            string obj_cast_fp16_to_fp32_dtype_0 = const()[name = string("obj_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [3]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14204672)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15023936)))];
+            tensor<fp16, [1, 1, 640]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = input_5_cast_fp16_0)[name = string("transpose_2")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [1]> f_axes_0 = const()[name = string("f_axes_0"), val = tensor<int32, [1]>([2])];
+            string encoder_proj_to_fp16_dtype_0 = const()[name = string("encoder_proj_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 1, 640]> encoder_proj_to_fp16 = cast(dtype = encoder_proj_to_fp16_dtype_0, x = encoder_proj)[name = string("cast_3")];
+            tensor<fp16, [1, 1, 1, 640]> f_cast_fp16 = expand_dims(axes = f_axes_0, x = encoder_proj_to_fp16)[name = string("f_cast_fp16")];
+            tensor<int32, [1]> g_axes_0 = const()[name = string("g_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> g_cast_fp16 = expand_dims(axes = g_axes_0, x = linear_0_cast_fp16)[name = string("g_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_9_cast_fp16 = add(x = f_cast_fp16, y = g_cast_fp16)[name = string("input_9_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_11_cast_fp16 = relu(x = input_9_cast_fp16)[name = string("input_11_cast_fp16")];
+            tensor<fp16, [849, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15025280)))];
+            tensor<fp16, [849]> joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [849]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16112064)))];
+            tensor<fp16, [1, 1, 1, 849]> linear_1_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_1_cast_fp16")];
+            int32 var_83 = const()[name = string("op_83"), val = int32(-1)];
+            tensor<fp16, [1, 1, 1, 849]> var_85_softmax_cast_fp16 = softmax(axis = var_83, x = linear_1_cast_fp16)[name = string("op_85_softmax_cast_fp16")];
+            fp32 var_85_epsilon_0 = const()[name = string("op_85_epsilon_0"), val = fp32(0x1p-149)];
+            tensor<fp16, [1, 1, 1, 849]> var_85_cast_fp16 = log(epsilon = var_85_epsilon_0, x = var_85_softmax_cast_fp16)[name = string("op_85_cast_fp16")];
+            string var_85_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_85_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 1, 1, 849]> logits = cast(dtype = var_85_cast_fp16_to_fp32_dtype_0, x = var_85_cast_fp16)[name = string("cast_2")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = string("cast_4")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = string("cast_5")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = string("token_length_tmp")];
+        } -> (logits, h_out, c_out);
+}
\ No newline at end of file
diff --git a/fr/560ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin b/fr/560ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..e37358fc0d446e079fe87526671dc411c4b37591
--- /dev/null
+++ b/fr/560ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c691d3004f2ac6a956208569a29d899a7f8e5030087d53fbd488785fa0e8efad
+size 16113826
diff --git a/fr/560ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/560ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..b701505b2faf5efd4a3c1024077f4834b1f94d37
--- /dev/null
+++ b/fr/560ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4196a821e716db75328fdd2f99f83214ecd9e6e6481a96a88907fe1c4232f143
+size 14630
diff --git a/fr/560ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/560ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..e37358fc0d446e079fe87526671dc411c4b37591
--- /dev/null
+++ b/fr/560ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c691d3004f2ac6a956208569a29d899a7f8e5030087d53fbd488785fa0e8efad
+size 16113826
diff --git a/fr/560ms/decoder_joint_noencproj.mlpackage/Manifest.json b/fr/560ms/decoder_joint_noencproj.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..4709121d3572643a899871aa076a46f4c02a7ef8
--- /dev/null
+++ b/fr/560ms/decoder_joint_noencproj.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "1D4938C1-F035-4D52-9376-C0FB72335508": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "BF28BA0A-268B-4C46-AB97-C0173D8B55F0": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "BF28BA0A-268B-4C46-AB97-C0173D8B55F0"
+}
diff --git a/fr/560ms/encoder.mlmodelc/analytics/coremldata.bin b/fr/560ms/encoder.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..f631009aa2b7c2a2f48501a3e27f90a0688e220e
--- /dev/null
+++ b/fr/560ms/encoder.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c7172b10d4d8cda6411e6356b4ba90d9d72d946859055983012fa4a9c9984deb
+size 243
diff --git a/fr/560ms/encoder.mlmodelc/coremldata.bin b/fr/560ms/encoder.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..73821303276998203e671e0d2f202f576644afb5
--- /dev/null
+++ b/fr/560ms/encoder.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:120f379cdd92608df7ba26ee2b8848c69fdee50cfc31bf6b59ae9765fecf74c5
+size 632
diff --git a/fr/560ms/encoder.mlmodelc/model.mil b/fr/560ms/encoder.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..d6f0179d3761ce7d81034c775fc7550c16626d14
--- /dev/null
+++ b/fr/560ms/encoder.mlmodelc/model.mil
@@ -0,0 +1,4434 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})]
+{
+    func main<ios18>(tensor<fp32, [1, 24, 42, 1024]> cache_channel, tensor<int32, [1]> cache_len, tensor<fp32, [1, 24, 1024, 8]> cache_time, tensor<fp32, [1, 128, 65]> mel, tensor<int32, [1]> mel_length, tensor<int32, [1]> prompt_id) {
+            tensor<int32, [4]> value_3_perm_0 = const()[name = string("value_3_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string cache_channel_to_fp16_dtype_0 = const()[name = string("cache_channel_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<int32, [4]> value_5_perm_0 = const()[name = string("value_5_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string cache_time_to_fp16_dtype_0 = const()[name = string("cache_time_to_fp16_dtype_0"), val = string("fp16")];
+            int32 var_59 = const()[name = string("op_59"), val = int32(-1)];
+            int32 var_68 = const()[name = string("op_68"), val = int32(1)];
+            tensor<int32, [3]> x_1_perm_0 = const()[name = string("x_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string mel_to_fp16_dtype_0 = const()[name = string("mel_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<int32, [1]> tensor_1_axes_0 = const()[name = string("tensor_1_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 128, 65]> mel_to_fp16 = cast(dtype = mel_to_fp16_dtype_0, x = mel)[name = string("cast_21")];
+            tensor<fp16, [1, 65, 128]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = mel_to_fp16)[name = string("transpose_367")];
+            tensor<fp16, [1, 1, 65, 128]> tensor_1_cast_fp16 = expand_dims(axes = tensor_1_axes_0, x = x_1_cast_fp16)[name = string("tensor_1_cast_fp16")];
+            tensor<int32, [1, 65]> expand_dims_0 = const()[name = string("expand_dims_0"), val = tensor<int32, [1, 65]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<int32, [1]> var_137_axes_0 = const()[name = string("op_137_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_137 = expand_dims(axes = var_137_axes_0, x = mel_length)[name = string("op_137")];
+            tensor<bool, [1, 65]> time_mask_1 = less(x = expand_dims_0, y = var_137)[name = string("time_mask_1")];
+            tensor<int32, [1]> var_139_axes_0 = const()[name = string("op_139_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 65, 1]> var_139 = expand_dims(axes = var_139_axes_0, x = time_mask_1)[name = string("op_139")];
+            tensor<int32, [3]> var_141_reps_0 = const()[name = string("op_141_reps_0"), val = tensor<int32, [3]>([1, 1, 128])];
+            tensor<bool, [1, 65, 128]> var_141 = tile(reps = var_141_reps_0, x = var_139)[name = string("op_141")];
+            tensor<int32, [1]> var_147_axes_0 = const()[name = string("op_147_axes_0"), val = tensor<int32, [1]>([1])];
+            string mask_1_to_fp16_dtype_0 = const()[name = string("mask_1_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 65, 128]> var_141_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_141)[name = string("cast_20")];
+            tensor<fp16, [1, 1, 65, 128]> var_147_cast_fp16 = expand_dims(axes = var_147_axes_0, x = var_141_to_fp16)[name = string("op_147_cast_fp16")];
+            tensor<fp16, [1, 1, 65, 128]> input_1_cast_fp16 = mul(x = tensor_1_cast_fp16, y = var_147_cast_fp16)[name = string("input_1_cast_fp16")];
+            tensor<int32, [8]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("constant")];
+            fp16 const_9_to_fp16 = const()[name = string("const_9_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 68, 131]> input_3_cast_fp16 = pad(constant_val = const_9_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")];
+            string tensor_3_pad_type_0 = const()[name = string("tensor_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_3_strides_0 = const()[name = string("tensor_3_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, [4]> tensor_3_pad_0 = const()[name = string("tensor_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_3_dilations_0 = const()[name = string("tensor_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 tensor_3_groups_0 = const()[name = string("tensor_3_groups_0"), val = int32(1)];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_0_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2816))))[name = string("encoder_pre_encode_conv_0_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_0_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_0_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3392)))];
+            tensor<fp16, [1, 256, 33, 65]> tensor_3_cast_fp16 = conv(bias = encoder_pre_encode_conv_0_bias_to_fp16, dilations = tensor_3_dilations_0, groups = tensor_3_groups_0, pad = tensor_3_pad_0, pad_type = tensor_3_pad_type_0, strides = tensor_3_strides_0, weight = encoder_pre_encode_conv_0_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("tensor_3_cast_fp16")];
+            string current_lengths_1_to_fp16_dtype_0 = const()[name = string("current_lengths_1_to_fp16_dtype_0"), val = string("fp16")];
+            fp16 var_160_promoted_to_fp16 = const()[name = string("op_160_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> mel_length_to_fp16 = cast(dtype = current_lengths_1_to_fp16_dtype_0, x = mel_length)[name = string("cast_19")];
+            tensor<fp16, [1]> var_161_cast_fp16 = add(x = mel_length_to_fp16, y = var_160_promoted_to_fp16)[name = string("op_161_cast_fp16")];
+            fp16 var_162_promoted_to_fp16 = const()[name = string("op_162_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> var_163_cast_fp16 = add(x = var_161_cast_fp16, y = var_162_promoted_to_fp16)[name = string("op_163_cast_fp16")];
+            fp16 var_164_promoted_to_fp16 = const()[name = string("op_164_promoted_to_fp16"), val = fp16(0x1.8p+1)];
+            tensor<fp16, [1]> var_165_cast_fp16 = sub(x = var_163_cast_fp16, y = var_164_promoted_to_fp16)[name = string("op_165_cast_fp16")];
+            fp16 var_56_promoted_to_fp16 = const()[name = string("op_56_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> floor_div_0_cast_fp16 = floor_div(x = var_165_cast_fp16, y = var_56_promoted_to_fp16)[name = string("floor_div_0_cast_fp16")];
+            fp16 var_167_promoted_to_fp16 = const()[name = string("op_167_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_3_cast_fp16 = add(x = floor_div_0_cast_fp16, y = var_167_promoted_to_fp16)[name = string("current_lengths_3_cast_fp16")];
+            string lengths_19_dtype_0 = const()[name = string("lengths_19_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 33]> expand_dims_1 = const()[name = string("expand_dims_1"), val = tensor<int32, [1, 33]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3968)))];
+            tensor<int32, [1]> var_176_axes_0 = const()[name = string("op_176_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_3_cast_fp16_to_int32 = cast(dtype = lengths_19_dtype_0, x = current_lengths_3_cast_fp16)[name = string("cast_18")];
+            tensor<int32, [1, 1]> var_176 = expand_dims(axes = var_176_axes_0, x = current_lengths_3_cast_fp16_to_int32)[name = string("op_176")];
+            tensor<bool, [1, 33]> time_mask_3 = less(x = expand_dims_1, y = var_176)[name = string("time_mask_3")];
+            tensor<int32, [1]> var_178_axes_0 = const()[name = string("op_178_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 33, 1]> var_178 = expand_dims(axes = var_178_axes_0, x = time_mask_3)[name = string("op_178")];
+            tensor<int32, [3]> var_180_reps_0 = const()[name = string("op_180_reps_0"), val = tensor<int32, [3]>([1, 1, 65])];
+            tensor<bool, [1, 33, 65]> var_180 = tile(reps = var_180_reps_0, x = var_178)[name = string("op_180")];
+            tensor<int32, [1]> var_186_axes_0 = const()[name = string("op_186_axes_0"), val = tensor<int32, [1]>([1])];
+            string mask_3_to_fp16_dtype_0 = const()[name = string("mask_3_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 33, 65]> var_180_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_180)[name = string("cast_17")];
+            tensor<fp16, [1, 1, 33, 65]> var_186_cast_fp16 = expand_dims(axes = var_186_axes_0, x = var_180_to_fp16)[name = string("op_186_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_3_reps_0 = const()[name = string("expanded_mask_3_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 33, 65]> expanded_mask_3_cast_fp16 = tile(reps = expanded_mask_3_reps_0, x = var_186_cast_fp16)[name = string("expanded_mask_3_cast_fp16")];
+            tensor<fp16, [1, 256, 33, 65]> input_5_cast_fp16 = mul(x = tensor_3_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<fp16, [1, 256, 33, 65]> tensor_5_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("tensor_5_cast_fp16")];
+            tensor<fp16, [1, 256, 33, 65]> input_7_cast_fp16 = mul(x = tensor_5_cast_fp16, y = expanded_mask_3_cast_fp16)[name = string("input_7_cast_fp16")];
+            tensor<int32, [8]> input_9_pad_0 = const()[name = string("input_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            string input_9_mode_0 = const()[name = string("input_9_mode_0"), val = string("constant")];
+            fp16 const_23_to_fp16 = const()[name = string("const_23_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 256, 36, 68]> input_9_cast_fp16 = pad(constant_val = const_23_to_fp16, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")];
+            string tensor_7_pad_type_0 = const()[name = string("tensor_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_7_strides_0 = const()[name = string("tensor_7_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            int32 tensor_7_groups_0 = const()[name = string("tensor_7_groups_0"), val = int32(256)];
+            tensor<int32, [4]> tensor_7_pad_0 = const()[name = string("tensor_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_7_dilations_0 = const()[name = string("tensor_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4224))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6592))))[name = string("encoder_pre_encode_conv_2_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_2_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_2_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7168)))];
+            tensor<fp16, [1, 256, 17, 33]> tensor_7_cast_fp16 = conv(bias = encoder_pre_encode_conv_2_bias_to_fp16, dilations = tensor_7_dilations_0, groups = tensor_7_groups_0, pad = tensor_7_pad_0, pad_type = tensor_7_pad_type_0, strides = tensor_7_strides_0, weight = encoder_pre_encode_conv_2_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("tensor_7_cast_fp16")];
+            fp16 var_208_promoted_to_fp16 = const()[name = string("op_208_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> var_209_cast_fp16 = add(x = current_lengths_3_cast_fp16, y = var_208_promoted_to_fp16)[name = string("op_209_cast_fp16")];
+            fp16 var_210_promoted_to_fp16 = const()[name = string("op_210_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> var_211_cast_fp16 = add(x = var_209_cast_fp16, y = var_210_promoted_to_fp16)[name = string("op_211_cast_fp16")];
+            fp16 var_212_promoted_to_fp16 = const()[name = string("op_212_promoted_to_fp16"), val = fp16(0x1.8p+1)];
+            tensor<fp16, [1]> var_213_cast_fp16 = sub(x = var_211_cast_fp16, y = var_212_promoted_to_fp16)[name = string("op_213_cast_fp16")];
+            fp16 var_56_promoted_1_to_fp16 = const()[name = string("op_56_promoted_1_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> floor_div_1_cast_fp16 = floor_div(x = var_213_cast_fp16, y = var_56_promoted_1_to_fp16)[name = string("floor_div_1_cast_fp16")];
+            fp16 var_215_promoted_to_fp16 = const()[name = string("op_215_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_5_cast_fp16 = add(x = floor_div_1_cast_fp16, y = var_215_promoted_to_fp16)[name = string("current_lengths_5_cast_fp16")];
+            string lengths_21_dtype_0 = const()[name = string("lengths_21_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 17]> expand_dims_2 = const()[name = string("expand_dims_2"), val = tensor<int32, [1, 17]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7744)))];
+            tensor<int32, [1]> var_224_axes_0 = const()[name = string("op_224_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_5_cast_fp16_to_int32 = cast(dtype = lengths_21_dtype_0, x = current_lengths_5_cast_fp16)[name = string("cast_16")];
+            tensor<int32, [1, 1]> var_224 = expand_dims(axes = var_224_axes_0, x = current_lengths_5_cast_fp16_to_int32)[name = string("op_224")];
+            tensor<bool, [1, 17]> time_mask_5 = less(x = expand_dims_2, y = var_224)[name = string("time_mask_5")];
+            tensor<int32, [1]> var_226_axes_0 = const()[name = string("op_226_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 17, 1]> var_226 = expand_dims(axes = var_226_axes_0, x = time_mask_5)[name = string("op_226")];
+            tensor<int32, [3]> var_228_reps_0 = const()[name = string("op_228_reps_0"), val = tensor<int32, [3]>([1, 1, 33])];
+            tensor<bool, [1, 17, 33]> var_228 = tile(reps = var_228_reps_0, x = var_226)[name = string("op_228")];
+            tensor<int32, [1]> var_234_axes_0 = const()[name = string("op_234_axes_0"), val = tensor<int32, [1]>([1])];
+            string mask_5_to_fp16_dtype_0 = const()[name = string("mask_5_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 17, 33]> var_228_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_228)[name = string("cast_15")];
+            tensor<fp16, [1, 1, 17, 33]> var_234_cast_fp16 = expand_dims(axes = var_234_axes_0, x = var_228_to_fp16)[name = string("op_234_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_7_reps_0 = const()[name = string("expanded_mask_7_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 17, 33]> expanded_mask_7_cast_fp16 = tile(reps = expanded_mask_7_reps_0, x = var_234_cast_fp16)[name = string("expanded_mask_7_cast_fp16")];
+            tensor<fp16, [1, 256, 17, 33]> input_11_cast_fp16 = mul(x = tensor_7_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_11_cast_fp16")];
+            string tensor_9_pad_type_0 = const()[name = string("tensor_9_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_9_strides_0 = const()[name = string("tensor_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> tensor_9_pad_0 = const()[name = string("tensor_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_9_dilations_0 = const()[name = string("tensor_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 tensor_9_groups_0 = const()[name = string("tensor_9_groups_0"), val = int32(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_3_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7936))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73536))))[name = string("encoder_pre_encode_conv_3_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_3_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_3_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74112)))];
+            tensor<fp16, [1, 256, 17, 33]> tensor_9_cast_fp16 = conv(bias = encoder_pre_encode_conv_3_bias_to_fp16, dilations = tensor_9_dilations_0, groups = tensor_9_groups_0, pad = tensor_9_pad_0, pad_type = tensor_9_pad_type_0, strides = tensor_9_strides_0, weight = encoder_pre_encode_conv_3_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("tensor_9_cast_fp16")];
+            tensor<fp16, [1, 256, 17, 33]> input_13_cast_fp16 = mul(x = tensor_9_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_13_cast_fp16")];
+            tensor<fp16, [1, 256, 17, 33]> tensor_11_cast_fp16 = relu(x = input_13_cast_fp16)[name = string("tensor_11_cast_fp16")];
+            tensor<fp16, [1, 256, 17, 33]> input_15_cast_fp16 = mul(x = tensor_11_cast_fp16, y = expanded_mask_7_cast_fp16)[name = string("input_15_cast_fp16")];
+            tensor<int32, [8]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            string input_17_mode_0 = const()[name = string("input_17_mode_0"), val = string("constant")];
+            fp16 const_41_to_fp16 = const()[name = string("const_41_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 256, 20, 36]> input_17_cast_fp16 = pad(constant_val = const_41_to_fp16, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")];
+            string tensor_13_pad_type_0 = const()[name = string("tensor_13_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_13_strides_0 = const()[name = string("tensor_13_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            int32 tensor_13_groups_0 = const()[name = string("tensor_13_groups_0"), val = int32(256)];
+            tensor<int32, [4]> tensor_13_pad_0 = const()[name = string("tensor_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_13_dilations_0 = const()[name = string("tensor_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp16, [256, 1, 3, 3]> encoder_pre_encode_conv_5_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(74688))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77056))))[name = string("encoder_pre_encode_conv_5_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_5_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_5_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77632)))];
+            tensor<fp16, [1, 256, 9, 17]> tensor_13_cast_fp16 = conv(bias = encoder_pre_encode_conv_5_bias_to_fp16, dilations = tensor_13_dilations_0, groups = tensor_13_groups_0, pad = tensor_13_pad_0, pad_type = tensor_13_pad_type_0, strides = tensor_13_strides_0, weight = encoder_pre_encode_conv_5_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("tensor_13_cast_fp16")];
+            fp16 var_271_promoted_to_fp16 = const()[name = string("op_271_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> var_272_cast_fp16 = add(x = current_lengths_5_cast_fp16, y = var_271_promoted_to_fp16)[name = string("op_272_cast_fp16")];
+            fp16 var_273_promoted_to_fp16 = const()[name = string("op_273_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> var_274_cast_fp16 = add(x = var_272_cast_fp16, y = var_273_promoted_to_fp16)[name = string("op_274_cast_fp16")];
+            fp16 var_275_promoted_to_fp16 = const()[name = string("op_275_promoted_to_fp16"), val = fp16(0x1.8p+1)];
+            tensor<fp16, [1]> var_276_cast_fp16 = sub(x = var_274_cast_fp16, y = var_275_promoted_to_fp16)[name = string("op_276_cast_fp16")];
+            fp16 var_56_promoted_2_to_fp16 = const()[name = string("op_56_promoted_2_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1]> floor_div_2_cast_fp16 = floor_div(x = var_276_cast_fp16, y = var_56_promoted_2_to_fp16)[name = string("floor_div_2_cast_fp16")];
+            fp16 var_278_promoted_to_fp16 = const()[name = string("op_278_promoted_to_fp16"), val = fp16(0x1p+0)];
+            tensor<fp16, [1]> current_lengths_cast_fp16 = add(x = floor_div_2_cast_fp16, y = var_278_promoted_to_fp16)[name = string("current_lengths_cast_fp16")];
+            string lengths_dtype_0 = const()[name = string("lengths_dtype_0"), val = string("int32")];
+            tensor<int32, [1, 9]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1, 9]>([[0, 1, 2, 3, 4, 5, 6, 7, 8]])];
+            tensor<int32, [1]> var_287_axes_0 = const()[name = string("op_287_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> current_lengths_cast_fp16_to_int32 = cast(dtype = lengths_dtype_0, x = current_lengths_cast_fp16)[name = string("cast_14")];
+            tensor<int32, [1, 1]> var_287 = expand_dims(axes = var_287_axes_0, x = current_lengths_cast_fp16_to_int32)[name = string("op_287")];
+            tensor<bool, [1, 9]> time_mask = less(x = expand_dims_3, y = var_287)[name = string("time_mask")];
+            tensor<int32, [1]> var_289_axes_0 = const()[name = string("op_289_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 9, 1]> var_289 = expand_dims(axes = var_289_axes_0, x = time_mask)[name = string("op_289")];
+            tensor<int32, [3]> var_291_reps_0 = const()[name = string("op_291_reps_0"), val = tensor<int32, [3]>([1, 1, 17])];
+            tensor<bool, [1, 9, 17]> var_291 = tile(reps = var_291_reps_0, x = var_289)[name = string("op_291")];
+            tensor<int32, [1]> var_297_axes_0 = const()[name = string("op_297_axes_0"), val = tensor<int32, [1]>([1])];
+            string mask_7_to_fp16_dtype_0 = const()[name = string("mask_7_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 9, 17]> var_291_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_291)[name = string("cast_13")];
+            tensor<fp16, [1, 1, 9, 17]> var_297_cast_fp16 = expand_dims(axes = var_297_axes_0, x = var_291_to_fp16)[name = string("op_297_cast_fp16")];
+            tensor<int32, [4]> expanded_mask_13_reps_0 = const()[name = string("expanded_mask_13_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp16, [1, 256, 9, 17]> expanded_mask_13_cast_fp16 = tile(reps = expanded_mask_13_reps_0, x = var_297_cast_fp16)[name = string("expanded_mask_13_cast_fp16")];
+            tensor<fp16, [1, 256, 9, 17]> input_19_cast_fp16 = mul(x = tensor_13_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_19_cast_fp16")];
+            string tensor_15_pad_type_0 = const()[name = string("tensor_15_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> tensor_15_strides_0 = const()[name = string("tensor_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> tensor_15_pad_0 = const()[name = string("tensor_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_15_dilations_0 = const()[name = string("tensor_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 tensor_15_groups_0 = const()[name = string("tensor_15_groups_0"), val = int32(1)];
+            tensor<fp16, [256, 256, 1, 1]> encoder_pre_encode_conv_6_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [256, 256, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78208))), scale = tensor<fp16, [256, 1, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143808))))[name = string("encoder_pre_encode_conv_6_weight_to_fp16_quantized")];
+            tensor<fp16, [256]> encoder_pre_encode_conv_6_bias_to_fp16 = const()[name = string("encoder_pre_encode_conv_6_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144384)))];
+            tensor<fp16, [1, 256, 9, 17]> tensor_15_cast_fp16 = conv(bias = encoder_pre_encode_conv_6_bias_to_fp16, dilations = tensor_15_dilations_0, groups = tensor_15_groups_0, pad = tensor_15_pad_0, pad_type = tensor_15_pad_type_0, strides = tensor_15_strides_0, weight = encoder_pre_encode_conv_6_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("tensor_15_cast_fp16")];
+            tensor<fp16, [1, 256, 9, 17]> input_21_cast_fp16 = mul(x = tensor_15_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("input_21_cast_fp16")];
+            tensor<fp16, [1, 256, 9, 17]> tensor_cast_fp16 = relu(x = input_21_cast_fp16)[name = string("tensor_cast_fp16")];
+            tensor<fp16, [1, 256, 9, 17]> x_3_cast_fp16 = mul(x = tensor_cast_fp16, y = expanded_mask_13_cast_fp16)[name = string("x_3_cast_fp16")];
+            tensor<int32, [4]> var_331_perm_0 = const()[name = string("op_331_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_332 = const()[name = string("op_332"), val = tensor<int32, [3]>([1, 9, -1])];
+            tensor<fp16, [1, 9, 256, 17]> var_331_cast_fp16 = transpose(perm = var_331_perm_0, x = x_3_cast_fp16)[name = string("transpose_366")];
+            tensor<fp16, [1, 9, 4352]> input_23_cast_fp16 = reshape(shape = var_332, x = var_331_cast_fp16)[name = string("input_23_cast_fp16")];
+            tensor<fp16, [1024, 4352]> encoder_pre_encode_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4352]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144960))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4601472))))[name = string("encoder_pre_encode_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_pre_encode_out_bias_to_fp16 = const()[name = string("encoder_pre_encode_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4603584)))];
+            tensor<fp16, [1, 9, 1024]> linear_0_cast_fp16 = linear(bias = encoder_pre_encode_out_bias_to_fp16, weight = encoder_pre_encode_out_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [3]> var_342_begin_0 = const()[name = string("op_342_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_342_end_0 = const()[name = string("op_342_end_0"), val = tensor<int32, [3]>([1, 9, 1024])];
+            tensor<bool, [3]> var_342_end_mask_0 = const()[name = string("op_342_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 7, 1024]> var_342_cast_fp16 = slice_by_index(begin = var_342_begin_0, end = var_342_end_0, end_mask = var_342_end_mask_0, x = linear_0_cast_fp16)[name = string("op_342_cast_fp16")];
+            int32 var_344 = const()[name = string("op_344"), val = int32(2)];
+            tensor<int32, [1]> var_345 = sub(x = current_lengths_cast_fp16_to_int32, y = var_344)[name = string("op_345")];
+            string var_345_promoted_to_fp16_dtype_0 = const()[name = string("op_345_promoted_to_fp16_dtype_0"), val = string("fp16")];
+            fp16 var_62_promoted_to_fp16 = const()[name = string("op_62_promoted_to_fp16"), val = fp16(0x0p+0)];
+            fp16 const_61_to_fp16 = const()[name = string("const_61_to_fp16"), val = fp16(inf)];
+            tensor<fp16, [1]> var_345_to_fp16 = cast(dtype = var_345_promoted_to_fp16_dtype_0, x = var_345)[name = string("cast_12")];
+            tensor<fp16, [1]> clip_0_cast_fp16 = clip(alpha = var_62_promoted_to_fp16, beta = const_61_to_fp16, x = var_345_to_fp16)[name = string("clip_0_cast_fp16")];
+            tensor<int32, [1]> max_audio_length_1 = const()[name = string("max_audio_length_1"), val = tensor<int32, [1]>([7])];
+            fp16 var_361_promoted_to_fp16 = const()[name = string("op_361_promoted_to_fp16"), val = fp16(0x1.5p+5)];
+            tensor<fp16, [1]> padding_length_cast_fp16 = add(x = clip_0_cast_fp16, y = var_361_promoted_to_fp16)[name = string("padding_length_cast_fp16")];
+            int32 const_63 = const()[name = string("const_63"), val = int32(-1)];
+            tensor<int32, [1]> var_363 = mul(x = cache_len, y = const_63)[name = string("op_363")];
+            int32 var_364 = const()[name = string("op_364"), val = int32(42)];
+            tensor<int32, [1]> offset = add(x = var_363, y = var_364)[name = string("offset")];
+            tensor<int32, [1]> var_404_axes_0 = const()[name = string("op_404_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1, 1]> var_404_cast_fp16 = expand_dims(axes = var_404_axes_0, x = padding_length_cast_fp16)[name = string("op_404_cast_fp16")];
+            tensor<fp16, [1, 49]> var_403_promoted_to_fp16 = const()[name = string("op_403_promoted_to_fp16"), val = tensor<fp16, [1, 49]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4605696)))];
+            tensor<bool, [1, 49]> pad_mask_1_cast_fp16 = less(x = var_403_promoted_to_fp16, y = var_404_cast_fp16)[name = string("pad_mask_1_cast_fp16")];
+            tensor<int32, [1, 49]> expand_dims_5 = const()[name = string("expand_dims_5"), val = tensor<int32, [1, 49]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4605888)))];
+            tensor<int32, [1]> var_410_axes_0 = const()[name = string("op_410_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 1]> var_410 = expand_dims(axes = var_410_axes_0, x = offset)[name = string("op_410")];
+            tensor<bool, [1, 49]> pad_mask_off = greater_equal(x = expand_dims_5, y = var_410)[name = string("pad_mask_off")];
+            tensor<bool, [1, 49]> pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1_cast_fp16)[name = string("pad_mask_3")];
+            tensor<int32, [1]> var_413_axes_0 = const()[name = string("op_413_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 49]> var_413 = expand_dims(axes = var_413_axes_0, x = pad_mask_3)[name = string("op_413")];
+            tensor<int32, [3]> var_414 = const()[name = string("op_414"), val = tensor<int32, [3]>([1, 49, 1])];
+            tensor<bool, [1, 49, 49]> pad_mask_for_att_mask_1 = tile(reps = var_414, x = var_413)[name = string("pad_mask_for_att_mask_1")];
+            tensor<int32, [3]> var_416_perm_0 = const()[name = string("op_416_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<bool, [1, 49, 49]> var_416 = transpose(perm = var_416_perm_0, x = pad_mask_for_att_mask_1)[name = string("transpose_365")];
+            tensor<bool, [1, 49, 49]> pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_416)[name = string("pad_mask_for_att_mask")];
+            tensor<bool, [1, 49, 49]> const_71 = const()[name = string("const_71"), val = tensor<bool, [1, 49, 49]>([[[true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])];
+            tensor<bool, [1, 49, 49]> att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_71)[name = string("att_mask_9")];
+            tensor<bool, [1, 49, 49]> att_mask = logical_not(x = att_mask_9)[name = string("att_mask")];
+            tensor<bool, [1, 49]> pad_mask_5 = logical_not(x = pad_mask_3)[name = string("pad_mask_5")];
+            tensor<int32, [2]> pad_mask_begin_0 = const()[name = string("pad_mask_begin_0"), val = tensor<int32, [2]>([0, 42])];
+            tensor<int32, [2]> pad_mask_end_0 = const()[name = string("pad_mask_end_0"), val = tensor<int32, [2]>([1, 49])];
+            tensor<bool, [2]> pad_mask_end_mask_0 = const()[name = string("pad_mask_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<bool, [1, 7]> pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = string("pad_mask")];
+            tensor<int32, [3]> mask_9_begin_0 = const()[name = string("mask_9_begin_0"), val = tensor<int32, [3]>([0, 42, 0])];
+            tensor<int32, [3]> mask_9_end_0 = const()[name = string("mask_9_end_0"), val = tensor<int32, [3]>([1, 49, 49])];
+            tensor<bool, [3]> mask_9_end_mask_0 = const()[name = string("mask_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<bool, [1, 7, 49]> mask_9 = slice_by_index(begin = mask_9_begin_0, end = mask_9_end_0, end_mask = mask_9_end_mask_0, x = att_mask)[name = string("mask_9")];
+            tensor<int32, [4]> cache_1_begin_0 = const()[name = string("cache_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_1_end_0 = const()[name = string("cache_1_end_0"), val = tensor<int32, [4]>([1, 1, 42, 1024])];
+            tensor<bool, [4]> cache_1_end_mask_0 = const()[name = string("cache_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_1_squeeze_mask_0 = const()[name = string("cache_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 24, 42, 1024]> cache_channel_to_fp16 = cast(dtype = cache_channel_to_fp16_dtype_0, x = cache_channel)[name = string("cast_11")];
+            tensor<fp16, [24, 1, 42, 1024]> value_3_cast_fp16 = transpose(perm = value_3_perm_0, x = cache_channel_to_fp16)[name = string("transpose_364")];
+            tensor<fp16, [1, 42, 1024]> cache_1_cast_fp16 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_1_cast_fp16")];
+            tensor<int32, [4]> cache_3_begin_0 = const()[name = string("cache_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_3_end_0 = const()[name = string("cache_3_end_0"), val = tensor<int32, [4]>([1, 1, 1024, 8])];
+            tensor<bool, [4]> cache_3_end_mask_0 = const()[name = string("cache_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_3_squeeze_mask_0 = const()[name = string("cache_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 24, 1024, 8]> cache_time_to_fp16 = cast(dtype = cache_time_to_fp16_dtype_0, x = cache_time)[name = string("cast_10")];
+            tensor<fp16, [24, 1, 1024, 8]> value_5_cast_fp16 = transpose(perm = value_5_perm_0, x = cache_time_to_fp16)[name = string("transpose_363")];
+            tensor<fp16, [1, 1024, 8]> cache_3_cast_fp16 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_3_cast_fp16")];
+            tensor<int32, [1]> input_27_axes_0 = const()[name = string("input_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4606208)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4608320)))];
+            fp16 var_42_to_fp16 = const()[name = string("op_42_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 7, 1024]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = encoder_layers_0_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_feed_forward1_weight_to_fp16, x = var_342_cast_fp16)[name = string("input_27_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4610432))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8804800))))[name = string("encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_0_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8813056)))];
+            tensor<fp16, [1, 7, 4096]> linear_1_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear1_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_31_cast_fp16 = silu(x = linear_1_cast_fp16)[name = string("input_31_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8821312))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13015680))))[name = string("encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13017792)))];
+            tensor<fp16, [1, 7, 1024]> linear_2_cast_fp16 = linear(bias = encoder_layers_0_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward1_linear2_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_2_cast_fp16")];
+            fp16 var_455_to_fp16 = const()[name = string("op_455_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_456_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_455_to_fp16)[name = string("op_456_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_37_cast_fp16 = add(x = var_342_cast_fp16, y = var_456_cast_fp16)[name = string("input_37_cast_fp16")];
+            tensor<int32, [1]> key_1_axes_0 = const()[name = string("key_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13019904)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13022016)))];
+            tensor<fp16, [1, 7, 1024]> key_1_cast_fp16 = layer_norm(axes = key_1_axes_0, beta = encoder_layers_0_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_self_att_weight_to_fp16, x = input_37_cast_fp16)[name = string("key_1_cast_fp16")];
+            bool input_39_interleave_0 = const()[name = string("input_39_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_39_cast_fp16 = concat(axis = var_68, interleave = input_39_interleave_0, values = (cache_1_cast_fp16, key_1_cast_fp16))[name = string("input_39_cast_fp16")];
+            tensor<int32, [3]> var_478_begin_0 = const()[name = string("op_478_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_478_end_0 = const()[name = string("op_478_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_478_end_mask_0 = const()[name = string("op_478_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_478_cast_fp16 = slice_by_index(begin = var_478_begin_0, end = var_478_end_0, end_mask = var_478_end_mask_0, x = cache_1_cast_fp16)[name = string("op_478_cast_fp16")];
+            bool var_484_interleave_0 = const()[name = string("op_484_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_484_cast_fp16 = concat(axis = var_68, interleave = var_484_interleave_0, values = (var_478_cast_fp16, key_1_cast_fp16))[name = string("op_484_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13024128))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14072768))))[name = string("encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14074880)))];
+            tensor<fp16, [1, 7, 1024]> linear_3_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_q_weight_to_fp16_quantized, x = key_1_cast_fp16)[name = string("linear_3_cast_fp16")];
+            tensor<int32, [4]> var_489 = const()[name = string("op_489"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_1_cast_fp16 = reshape(shape = var_489, x = linear_3_cast_fp16)[name = string("q_1_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14076992))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15125632))))[name = string("encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15127744)))];
+            tensor<fp16, [1, 49, 1024]> linear_4_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_k_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_4_cast_fp16")];
+            tensor<int32, [4]> var_494 = const()[name = string("op_494"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_1_cast_fp16 = reshape(shape = var_494, x = linear_4_cast_fp16)[name = string("k_1_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15129856))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16178496))))[name = string("encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16180608)))];
+            tensor<fp16, [1, 49, 1024]> linear_5_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_v_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_5_cast_fp16")];
+            tensor<int32, [4]> var_499 = const()[name = string("op_499"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_1_cast_fp16 = reshape(shape = var_499, x = linear_5_cast_fp16)[name = string("v_1_cast_fp16")];
+            tensor<int32, [4]> value_9_perm_0 = const()[name = string("value_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16182720)))];
+            tensor<fp16, [1, 7, 8, 128]> var_512_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_u_to_fp16)[name = string("op_512_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16184832)))];
+            tensor<fp16, [1, 7, 8, 128]> var_514_cast_fp16 = add(x = q_1_cast_fp16, y = encoder_layers_0_self_attn_pos_bias_v_to_fp16)[name = string("op_514_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_1_perm_0 = const()[name = string("q_with_bias_v_1_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_7_transpose_x_0 = const()[name = string("x_7_transpose_x_0"), val = bool(false)];
+            bool x_7_transpose_y_0 = const()[name = string("x_7_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_516_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16186944))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16286336))))[name = string("op_516_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_514_cast_fp16)[name = string("transpose_362")];
+            tensor<fp16, [1, 8, 7, 97]> x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = op_516_to_fp16_quantized)[name = string("x_7_cast_fp16")];
+            tensor<int32, [8]> x_9_pad_0 = const()[name = string("x_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_9_mode_0 = const()[name = string("x_9_mode_0"), val = string("constant")];
+            fp16 const_79_to_fp16 = const()[name = string("const_79_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_9_cast_fp16 = pad(constant_val = const_79_to_fp16, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7_cast_fp16)[name = string("x_9_cast_fp16")];
+            tensor<int32, [4]> var_524 = const()[name = string("op_524"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_11_cast_fp16 = reshape(shape = var_524, x = x_9_cast_fp16)[name = string("x_11_cast_fp16")];
+            tensor<int32, [4]> var_528_begin_0 = const()[name = string("op_528_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_528_end_0 = const()[name = string("op_528_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_528_end_mask_0 = const()[name = string("op_528_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_528_cast_fp16 = slice_by_index(begin = var_528_begin_0, end = var_528_end_0, end_mask = var_528_end_mask_0, x = x_11_cast_fp16)[name = string("op_528_cast_fp16")];
+            tensor<int32, [4]> var_529 = const()[name = string("op_529"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_1_cast_fp16 = reshape(shape = var_529, x = var_528_cast_fp16)[name = string("matrix_bd_1_cast_fp16")];
+            bool matrix_ac_1_transpose_x_0 = const()[name = string("matrix_ac_1_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_1_transpose_y_0 = const()[name = string("matrix_ac_1_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_97 = transpose(perm = transpose_97_perm_0, x = k_1_cast_fp16)[name = string("transpose_360")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_96 = transpose(perm = transpose_96_perm_0, x = var_512_cast_fp16)[name = string("transpose_361")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_1_cast_fp16 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_96, y = transpose_97)[name = string("matrix_ac_1_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_3_begin_0 = const()[name = string("matrix_bd_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_3_end_0 = const()[name = string("matrix_bd_3_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_3_end_mask_0 = const()[name = string("matrix_bd_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_3_cast_fp16 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1_cast_fp16)[name = string("matrix_bd_3_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_538_cast_fp16 = add(x = matrix_ac_1_cast_fp16, y = matrix_bd_3_cast_fp16)[name = string("op_538_cast_fp16")];
+            fp16 _inversed_scores_1_y_0_to_fp16 = const()[name = string("_inversed_scores_1_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_1_cast_fp16 = mul(x = var_538_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = string("_inversed_scores_1_cast_fp16")];
+            tensor<int32, [1]> mask_11_axes_0 = const()[name = string("mask_11_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 7, 49]> mask_11 = expand_dims(axes = mask_11_axes_0, x = mask_9)[name = string("mask_11")];
+            fp16 var_45_to_fp16 = const()[name = string("op_45_to_fp16"), val = fp16(-0x1.388p+13)];
+            tensor<fp16, [1, 8, 7, 49]> scores_3_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_1_cast_fp16, cond = mask_11)[name = string("scores_3_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_544_cast_fp16 = softmax(axis = var_59, x = scores_3_cast_fp16)[name = string("op_544_cast_fp16")];
+            fp16 var_44_to_fp16 = const()[name = string("op_44_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 49]> input_41_cast_fp16 = select(a = var_44_to_fp16, b = var_544_cast_fp16, cond = mask_11)[name = string("input_41_cast_fp16")];
+            bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)];
+            bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_9_cast_fp16 = transpose(perm = value_9_perm_0, x = v_1_cast_fp16)[name = string("transpose_359")];
+            tensor<fp16, [1, 8, 7, 128]> x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41_cast_fp16, y = value_9_cast_fp16)[name = string("x_13_cast_fp16")];
+            tensor<int32, [4]> var_548_perm_0 = const()[name = string("op_548_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_549 = const()[name = string("op_549"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_548_cast_fp16 = transpose(perm = var_548_perm_0, x = x_13_cast_fp16)[name = string("transpose_358")];
+            tensor<fp16, [1, 7, 1024]> input_43_cast_fp16 = reshape(shape = var_549, x = var_548_cast_fp16)[name = string("input_43_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16286656))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17335296))))[name = string("encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_0_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17337408)))];
+            tensor<fp16, [1, 7, 1024]> linear_7_cast_fp16 = linear(bias = encoder_layers_0_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_0_self_attn_linear_out_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_7_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_47_cast_fp16 = add(x = input_37_cast_fp16, y = linear_7_cast_fp16)[name = string("input_47_cast_fp16")];
+            tensor<int32, [1]> x_17_axes_0 = const()[name = string("x_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17339520)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17341632)))];
+            tensor<fp16, [1, 7, 1024]> x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = encoder_layers_0_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_conv_weight_to_fp16, x = input_47_cast_fp16)[name = string("x_17_cast_fp16")];
+            tensor<int32, [3]> input_49_perm_0 = const()[name = string("input_49_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_51_pad_type_0 = const()[name = string("input_51_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_51_strides_0 = const()[name = string("input_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_51_pad_0 = const()[name = string("input_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_51_dilations_0 = const()[name = string("input_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_51_groups_0 = const()[name = string("input_51_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17343744))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19440960))))[name = string("encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_49_cast_fp16 = transpose(perm = input_49_perm_0, x = x_17_cast_fp16)[name = string("transpose_357")];
+            tensor<fp16, [1, 2048, 7]> input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("input_51_cast_fp16")];
+            int32 x_19_split_num_splits_0 = const()[name = string("x_19_split_num_splits_0"), val = int32(2)];
+            int32 x_19_split_axis_0 = const()[name = string("x_19_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_19_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_19_split_cast_fp16_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51_cast_fp16)[name = string("x_19_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_19_split_1_sigmoid_cast_fp16 = sigmoid(x = x_19_split_cast_fp16_1)[name = string("x_19_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_19_cast_fp16 = mul(x = x_19_split_cast_fp16_0, y = x_19_split_1_sigmoid_cast_fp16)[name = string("x_19_cast_fp16")];
+            tensor<int32, [1]> var_575_axes_0 = const()[name = string("op_575_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 7]> var_575 = expand_dims(axes = var_575_axes_0, x = pad_mask)[name = string("op_575")];
+            tensor<fp16, [1, 1024, 7]> input_53_cast_fp16 = select(a = var_44_to_fp16, b = x_19_cast_fp16, cond = var_575)[name = string("input_53_cast_fp16")];
+            bool new_x_3_interleave_0 = const()[name = string("new_x_3_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_3_cast_fp16 = concat(axis = var_59, interleave = new_x_3_interleave_0, values = (cache_3_cast_fp16, input_53_cast_fp16))[name = string("new_x_3_cast_fp16")];
+            tensor<int32, [3]> var_588_begin_0 = const()[name = string("op_588_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_588_end_0 = const()[name = string("op_588_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_588_end_mask_0 = const()[name = string("op_588_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_588_cast_fp16 = slice_by_index(begin = var_588_begin_0, end = var_588_end_0, end_mask = var_588_end_mask_0, x = new_x_3_cast_fp16)[name = string("op_588_cast_fp16")];
+            string x_21_pad_type_0 = const()[name = string("x_21_pad_type_0"), val = string("valid")];
+            int32 x_21_groups_0 = const()[name = string("x_21_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_21_strides_0 = const()[name = string("x_21_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_21_pad_0 = const()[name = string("x_21_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_21_dilations_0 = const()[name = string("x_21_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19445120))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19454400))))[name = string("encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = encoder_layers_0_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_3_cast_fp16)[name = string("x_21_cast_fp16")];
+            tensor<int32, [3]> input_55_perm_0 = const()[name = string("input_55_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_23_axes_0 = const()[name = string("x_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19456512)))];
+            tensor<fp16, [1024]> encoder_layers_0_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_0_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19458624)))];
+            tensor<fp16, [1, 7, 1024]> input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = string("transpose_356")];
+            tensor<fp16, [1, 7, 1024]> x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = encoder_layers_0_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_conv_batch_norm_weight_to_fp16, x = input_55_cast_fp16)[name = string("x_23_cast_fp16")];
+            tensor<int32, [3]> input_57_perm_0 = const()[name = string("input_57_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_57_cast_fp16 = transpose(perm = input_57_perm_0, x = x_23_cast_fp16)[name = string("transpose_355")];
+            tensor<fp16, [1, 1024, 7]> input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = string("input_59_cast_fp16")];
+            string x_25_pad_type_0 = const()[name = string("x_25_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_25_strides_0 = const()[name = string("x_25_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_25_pad_0 = const()[name = string("x_25_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_25_dilations_0 = const()[name = string("x_25_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_25_groups_0 = const()[name = string("x_25_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19460736))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20509376))))[name = string("encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_25_cast_fp16 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("x_25_cast_fp16")];
+            tensor<int32, [3]> input_61_perm_0 = const()[name = string("input_61_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_61_cast_fp16 = transpose(perm = input_61_perm_0, x = x_25_cast_fp16)[name = string("transpose_354")];
+            tensor<fp16, [1, 7, 1024]> input_63_cast_fp16 = add(x = input_47_cast_fp16, y = input_61_cast_fp16)[name = string("input_63_cast_fp16")];
+            tensor<int32, [1]> input_65_axes_0 = const()[name = string("input_65_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20511488)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20513600)))];
+            tensor<fp16, [1, 7, 1024]> input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = encoder_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_feed_forward2_weight_to_fp16, x = input_63_cast_fp16)[name = string("input_65_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20515712))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24710080))))[name = string("encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_0_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24718336)))];
+            tensor<fp16, [1, 7, 4096]> linear_8_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear1_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_8_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_69_cast_fp16 = silu(x = linear_8_cast_fp16)[name = string("input_69_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24726592))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28920960))))[name = string("encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_0_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_0_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28923072)))];
+            tensor<fp16, [1, 7, 1024]> linear_9_cast_fp16 = linear(bias = encoder_layers_0_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_0_feed_forward2_linear2_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_9_cast_fp16")];
+            fp16 var_631_to_fp16 = const()[name = string("op_631_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_632_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_631_to_fp16)[name = string("op_632_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_75_cast_fp16 = add(x = input_63_cast_fp16, y = var_632_cast_fp16)[name = string("input_75_cast_fp16")];
+            tensor<int32, [1]> input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_0_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_0_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28925184)))];
+            tensor<fp16, [1024]> encoder_layers_0_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_0_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28927296)))];
+            tensor<fp16, [1, 7, 1024]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_layers_0_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_0_norm_out_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")];
+            tensor<int32, [4]> cache_5_begin_0 = const()[name = string("cache_5_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_5_end_0 = const()[name = string("cache_5_end_0"), val = tensor<int32, [4]>([2, 1, 42, 1024])];
+            tensor<bool, [4]> cache_5_end_mask_0 = const()[name = string("cache_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_5_squeeze_mask_0 = const()[name = string("cache_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_5_cast_fp16 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_5_cast_fp16")];
+            tensor<int32, [4]> cache_7_begin_0 = const()[name = string("cache_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_7_end_0 = const()[name = string("cache_7_end_0"), val = tensor<int32, [4]>([2, 1, 1024, 8])];
+            tensor<bool, [4]> cache_7_end_mask_0 = const()[name = string("cache_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_7_squeeze_mask_0 = const()[name = string("cache_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_7_cast_fp16 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_7_cast_fp16")];
+            tensor<int32, [1]> input_79_axes_0 = const()[name = string("input_79_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28929408)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28931520)))];
+            tensor<fp16, [1, 7, 1024]> input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = encoder_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_feed_forward1_weight_to_fp16, x = input_77_cast_fp16)[name = string("input_79_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28933632))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33128000))))[name = string("encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_1_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33136256)))];
+            tensor<fp16, [1, 7, 4096]> linear_10_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear1_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_10_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_83_cast_fp16 = silu(x = linear_10_cast_fp16)[name = string("input_83_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33144512))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37338880))))[name = string("encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37340992)))];
+            tensor<fp16, [1, 7, 1024]> linear_11_cast_fp16 = linear(bias = encoder_layers_1_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward1_linear2_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_11_cast_fp16")];
+            fp16 var_668_to_fp16 = const()[name = string("op_668_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_669_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_668_to_fp16)[name = string("op_669_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_89_cast_fp16 = add(x = input_77_cast_fp16, y = var_669_cast_fp16)[name = string("input_89_cast_fp16")];
+            tensor<int32, [1]> key_3_axes_0 = const()[name = string("key_3_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37343104)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37345216)))];
+            tensor<fp16, [1, 7, 1024]> key_3_cast_fp16 = layer_norm(axes = key_3_axes_0, beta = encoder_layers_1_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_self_att_weight_to_fp16, x = input_89_cast_fp16)[name = string("key_3_cast_fp16")];
+            bool input_91_interleave_0 = const()[name = string("input_91_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_91_cast_fp16 = concat(axis = var_68, interleave = input_91_interleave_0, values = (cache_5_cast_fp16, key_3_cast_fp16))[name = string("input_91_cast_fp16")];
+            tensor<int32, [3]> var_691_begin_0 = const()[name = string("op_691_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_691_end_0 = const()[name = string("op_691_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_691_end_mask_0 = const()[name = string("op_691_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_691_cast_fp16 = slice_by_index(begin = var_691_begin_0, end = var_691_end_0, end_mask = var_691_end_mask_0, x = cache_5_cast_fp16)[name = string("op_691_cast_fp16")];
+            bool var_697_interleave_0 = const()[name = string("op_697_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_697_cast_fp16 = concat(axis = var_68, interleave = var_697_interleave_0, values = (var_691_cast_fp16, key_3_cast_fp16))[name = string("op_697_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37347328))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38395968))))[name = string("encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38398080)))];
+            tensor<fp16, [1, 7, 1024]> linear_12_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_q_weight_to_fp16_quantized, x = key_3_cast_fp16)[name = string("linear_12_cast_fp16")];
+            tensor<int32, [4]> var_702 = const()[name = string("op_702"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_7_cast_fp16 = reshape(shape = var_702, x = linear_12_cast_fp16)[name = string("q_7_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38400192))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39448832))))[name = string("encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39450944)))];
+            tensor<fp16, [1, 49, 1024]> linear_13_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_k_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_13_cast_fp16")];
+            tensor<int32, [4]> var_707 = const()[name = string("op_707"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_5_cast_fp16 = reshape(shape = var_707, x = linear_13_cast_fp16)[name = string("k_5_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39453056))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40501696))))[name = string("encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40503808)))];
+            tensor<fp16, [1, 49, 1024]> linear_14_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_v_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_14_cast_fp16")];
+            tensor<int32, [4]> var_712 = const()[name = string("op_712"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_3_cast_fp16 = reshape(shape = var_712, x = linear_14_cast_fp16)[name = string("v_3_cast_fp16")];
+            tensor<int32, [4]> value_11_perm_0 = const()[name = string("value_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40505920)))];
+            tensor<fp16, [1, 7, 8, 128]> var_725_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_u_to_fp16)[name = string("op_725_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40508032)))];
+            tensor<fp16, [1, 7, 8, 128]> var_727_cast_fp16 = add(x = q_7_cast_fp16, y = encoder_layers_1_self_attn_pos_bias_v_to_fp16)[name = string("op_727_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_3_perm_0 = const()[name = string("q_with_bias_v_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_33_transpose_x_0 = const()[name = string("x_33_transpose_x_0"), val = bool(false)];
+            bool x_33_transpose_y_0 = const()[name = string("x_33_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_729_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40510144))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40609536))))[name = string("op_729_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_727_cast_fp16)[name = string("transpose_353")];
+            tensor<fp16, [1, 8, 7, 97]> x_33_cast_fp16 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = op_729_to_fp16_quantized)[name = string("x_33_cast_fp16")];
+            tensor<int32, [8]> x_35_pad_0 = const()[name = string("x_35_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_35_mode_0 = const()[name = string("x_35_mode_0"), val = string("constant")];
+            fp16 const_92_to_fp16 = const()[name = string("const_92_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_35_cast_fp16 = pad(constant_val = const_92_to_fp16, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33_cast_fp16)[name = string("x_35_cast_fp16")];
+            tensor<int32, [4]> var_737 = const()[name = string("op_737"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_37_cast_fp16 = reshape(shape = var_737, x = x_35_cast_fp16)[name = string("x_37_cast_fp16")];
+            tensor<int32, [4]> var_741_begin_0 = const()[name = string("op_741_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_741_end_0 = const()[name = string("op_741_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_741_end_mask_0 = const()[name = string("op_741_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_741_cast_fp16 = slice_by_index(begin = var_741_begin_0, end = var_741_end_0, end_mask = var_741_end_mask_0, x = x_37_cast_fp16)[name = string("op_741_cast_fp16")];
+            tensor<int32, [4]> var_742 = const()[name = string("op_742"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_5_cast_fp16 = reshape(shape = var_742, x = var_741_cast_fp16)[name = string("matrix_bd_5_cast_fp16")];
+            bool matrix_ac_3_transpose_x_0 = const()[name = string("matrix_ac_3_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_3_transpose_y_0 = const()[name = string("matrix_ac_3_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_99 = transpose(perm = transpose_99_perm_0, x = k_5_cast_fp16)[name = string("transpose_351")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_98 = transpose(perm = transpose_98_perm_0, x = var_725_cast_fp16)[name = string("transpose_352")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_3_cast_fp16 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_98, y = transpose_99)[name = string("matrix_ac_3_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_7_begin_0 = const()[name = string("matrix_bd_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_7_end_0 = const()[name = string("matrix_bd_7_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_7_end_mask_0 = const()[name = string("matrix_bd_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_7_cast_fp16 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5_cast_fp16)[name = string("matrix_bd_7_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_751_cast_fp16 = add(x = matrix_ac_3_cast_fp16, y = matrix_bd_7_cast_fp16)[name = string("op_751_cast_fp16")];
+            fp16 _inversed_scores_5_y_0_to_fp16 = const()[name = string("_inversed_scores_5_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_5_cast_fp16 = mul(x = var_751_cast_fp16, y = _inversed_scores_5_y_0_to_fp16)[name = string("_inversed_scores_5_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_7_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_5_cast_fp16, cond = mask_11)[name = string("scores_7_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_757_cast_fp16 = softmax(axis = var_59, x = scores_7_cast_fp16)[name = string("op_757_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_93_cast_fp16 = select(a = var_44_to_fp16, b = var_757_cast_fp16, cond = mask_11)[name = string("input_93_cast_fp16")];
+            bool x_39_transpose_x_0 = const()[name = string("x_39_transpose_x_0"), val = bool(false)];
+            bool x_39_transpose_y_0 = const()[name = string("x_39_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_11_cast_fp16 = transpose(perm = value_11_perm_0, x = v_3_cast_fp16)[name = string("transpose_350")];
+            tensor<fp16, [1, 8, 7, 128]> x_39_cast_fp16 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93_cast_fp16, y = value_11_cast_fp16)[name = string("x_39_cast_fp16")];
+            tensor<int32, [4]> var_761_perm_0 = const()[name = string("op_761_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_762 = const()[name = string("op_762"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_761_cast_fp16 = transpose(perm = var_761_perm_0, x = x_39_cast_fp16)[name = string("transpose_349")];
+            tensor<fp16, [1, 7, 1024]> input_95_cast_fp16 = reshape(shape = var_762, x = var_761_cast_fp16)[name = string("input_95_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40609856))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41658496))))[name = string("encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_1_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41660608)))];
+            tensor<fp16, [1, 7, 1024]> linear_16_cast_fp16 = linear(bias = encoder_layers_1_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_1_self_attn_linear_out_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_16_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_99_cast_fp16 = add(x = input_89_cast_fp16, y = linear_16_cast_fp16)[name = string("input_99_cast_fp16")];
+            tensor<int32, [1]> x_43_axes_0 = const()[name = string("x_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41662720)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41664832)))];
+            tensor<fp16, [1, 7, 1024]> x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = encoder_layers_1_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_conv_weight_to_fp16, x = input_99_cast_fp16)[name = string("x_43_cast_fp16")];
+            tensor<int32, [3]> input_101_perm_0 = const()[name = string("input_101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_103_pad_type_0 = const()[name = string("input_103_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_103_strides_0 = const()[name = string("input_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_103_pad_0 = const()[name = string("input_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_103_dilations_0 = const()[name = string("input_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_103_groups_0 = const()[name = string("input_103_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41666944))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43764160))))[name = string("encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_101_cast_fp16 = transpose(perm = input_101_perm_0, x = x_43_cast_fp16)[name = string("transpose_348")];
+            tensor<fp16, [1, 2048, 7]> input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = encoder_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("input_103_cast_fp16")];
+            int32 x_45_split_num_splits_0 = const()[name = string("x_45_split_num_splits_0"), val = int32(2)];
+            int32 x_45_split_axis_0 = const()[name = string("x_45_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_45_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_45_split_cast_fp16_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103_cast_fp16)[name = string("x_45_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_45_split_1_sigmoid_cast_fp16 = sigmoid(x = x_45_split_cast_fp16_1)[name = string("x_45_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_45_cast_fp16 = mul(x = x_45_split_cast_fp16_0, y = x_45_split_1_sigmoid_cast_fp16)[name = string("x_45_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_105_cast_fp16 = select(a = var_44_to_fp16, b = x_45_cast_fp16, cond = var_575)[name = string("input_105_cast_fp16")];
+            bool new_x_7_interleave_0 = const()[name = string("new_x_7_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_7_cast_fp16 = concat(axis = var_59, interleave = new_x_7_interleave_0, values = (cache_7_cast_fp16, input_105_cast_fp16))[name = string("new_x_7_cast_fp16")];
+            tensor<int32, [3]> var_801_begin_0 = const()[name = string("op_801_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_801_end_0 = const()[name = string("op_801_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_801_end_mask_0 = const()[name = string("op_801_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_801_cast_fp16 = slice_by_index(begin = var_801_begin_0, end = var_801_end_0, end_mask = var_801_end_mask_0, x = new_x_7_cast_fp16)[name = string("op_801_cast_fp16")];
+            string x_47_pad_type_0 = const()[name = string("x_47_pad_type_0"), val = string("valid")];
+            int32 x_47_groups_0 = const()[name = string("x_47_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_47_strides_0 = const()[name = string("x_47_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_47_pad_0 = const()[name = string("x_47_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_47_dilations_0 = const()[name = string("x_47_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43768320))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43777600))))[name = string("encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = encoder_layers_1_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_7_cast_fp16)[name = string("x_47_cast_fp16")];
+            tensor<int32, [3]> input_107_perm_0 = const()[name = string("input_107_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_49_axes_0 = const()[name = string("x_49_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43779712)))];
+            tensor<fp16, [1024]> encoder_layers_1_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_1_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43781824)))];
+            tensor<fp16, [1, 7, 1024]> input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_47_cast_fp16)[name = string("transpose_347")];
+            tensor<fp16, [1, 7, 1024]> x_49_cast_fp16 = layer_norm(axes = x_49_axes_0, beta = encoder_layers_1_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_conv_batch_norm_weight_to_fp16, x = input_107_cast_fp16)[name = string("x_49_cast_fp16")];
+            tensor<int32, [3]> input_109_perm_0 = const()[name = string("input_109_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_109_cast_fp16 = transpose(perm = input_109_perm_0, x = x_49_cast_fp16)[name = string("transpose_346")];
+            tensor<fp16, [1, 1024, 7]> input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = string("input_111_cast_fp16")];
+            string x_51_pad_type_0 = const()[name = string("x_51_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_51_strides_0 = const()[name = string("x_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_51_pad_0 = const()[name = string("x_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_51_dilations_0 = const()[name = string("x_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_51_groups_0 = const()[name = string("x_51_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43783936))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44832576))))[name = string("encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_51_cast_fp16 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = encoder_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("x_51_cast_fp16")];
+            tensor<int32, [3]> input_113_perm_0 = const()[name = string("input_113_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_113_cast_fp16 = transpose(perm = input_113_perm_0, x = x_51_cast_fp16)[name = string("transpose_345")];
+            tensor<fp16, [1, 7, 1024]> input_115_cast_fp16 = add(x = input_99_cast_fp16, y = input_113_cast_fp16)[name = string("input_115_cast_fp16")];
+            tensor<int32, [1]> input_117_axes_0 = const()[name = string("input_117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44834688)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44836800)))];
+            tensor<fp16, [1, 7, 1024]> input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = encoder_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_feed_forward2_weight_to_fp16, x = input_115_cast_fp16)[name = string("input_117_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44838912))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49033280))))[name = string("encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_1_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49041536)))];
+            tensor<fp16, [1, 7, 4096]> linear_17_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear1_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_17_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_121_cast_fp16 = silu(x = linear_17_cast_fp16)[name = string("input_121_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49049792))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53244160))))[name = string("encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_1_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_1_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53246272)))];
+            tensor<fp16, [1, 7, 1024]> linear_18_cast_fp16 = linear(bias = encoder_layers_1_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_1_feed_forward2_linear2_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_18_cast_fp16")];
+            fp16 var_844_to_fp16 = const()[name = string("op_844_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_845_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_844_to_fp16)[name = string("op_845_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_127_cast_fp16 = add(x = input_115_cast_fp16, y = var_845_cast_fp16)[name = string("input_127_cast_fp16")];
+            tensor<int32, [1]> input_129_axes_0 = const()[name = string("input_129_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_1_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_1_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53248384)))];
+            tensor<fp16, [1024]> encoder_layers_1_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_1_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53250496)))];
+            tensor<fp16, [1, 7, 1024]> input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = encoder_layers_1_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_1_norm_out_weight_to_fp16, x = input_127_cast_fp16)[name = string("input_129_cast_fp16")];
+            tensor<int32, [4]> cache_9_begin_0 = const()[name = string("cache_9_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_9_end_0 = const()[name = string("cache_9_end_0"), val = tensor<int32, [4]>([3, 1, 42, 1024])];
+            tensor<bool, [4]> cache_9_end_mask_0 = const()[name = string("cache_9_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_9_squeeze_mask_0 = const()[name = string("cache_9_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_9_cast_fp16 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_9_cast_fp16")];
+            tensor<int32, [4]> cache_11_begin_0 = const()[name = string("cache_11_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_11_end_0 = const()[name = string("cache_11_end_0"), val = tensor<int32, [4]>([3, 1, 1024, 8])];
+            tensor<bool, [4]> cache_11_end_mask_0 = const()[name = string("cache_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_11_squeeze_mask_0 = const()[name = string("cache_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_11_cast_fp16 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_11_cast_fp16")];
+            tensor<int32, [1]> input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53252608)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53254720)))];
+            tensor<fp16, [1, 7, 1024]> input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = encoder_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_feed_forward1_weight_to_fp16, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(53256832))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57451200))))[name = string("encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_2_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57459456)))];
+            tensor<fp16, [1, 7, 4096]> linear_19_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear1_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_19_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_135_cast_fp16 = silu(x = linear_19_cast_fp16)[name = string("input_135_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57467712))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61662080))))[name = string("encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61664192)))];
+            tensor<fp16, [1, 7, 1024]> linear_20_cast_fp16 = linear(bias = encoder_layers_2_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward1_linear2_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_20_cast_fp16")];
+            fp16 var_881_to_fp16 = const()[name = string("op_881_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_882_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_881_to_fp16)[name = string("op_882_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_141_cast_fp16 = add(x = input_129_cast_fp16, y = var_882_cast_fp16)[name = string("input_141_cast_fp16")];
+            tensor<int32, [1]> key_5_axes_0 = const()[name = string("key_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61666304)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61668416)))];
+            tensor<fp16, [1, 7, 1024]> key_5_cast_fp16 = layer_norm(axes = key_5_axes_0, beta = encoder_layers_2_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_self_att_weight_to_fp16, x = input_141_cast_fp16)[name = string("key_5_cast_fp16")];
+            bool input_143_interleave_0 = const()[name = string("input_143_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_143_cast_fp16 = concat(axis = var_68, interleave = input_143_interleave_0, values = (cache_9_cast_fp16, key_5_cast_fp16))[name = string("input_143_cast_fp16")];
+            tensor<int32, [3]> var_904_begin_0 = const()[name = string("op_904_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_904_end_0 = const()[name = string("op_904_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_904_end_mask_0 = const()[name = string("op_904_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_904_cast_fp16 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = cache_9_cast_fp16)[name = string("op_904_cast_fp16")];
+            bool var_910_interleave_0 = const()[name = string("op_910_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_910_cast_fp16 = concat(axis = var_68, interleave = var_910_interleave_0, values = (var_904_cast_fp16, key_5_cast_fp16))[name = string("op_910_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(61670528))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62719168))))[name = string("encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62721280)))];
+            tensor<fp16, [1, 7, 1024]> linear_21_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_q_weight_to_fp16_quantized, x = key_5_cast_fp16)[name = string("linear_21_cast_fp16")];
+            tensor<int32, [4]> var_915 = const()[name = string("op_915"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_13_cast_fp16 = reshape(shape = var_915, x = linear_21_cast_fp16)[name = string("q_13_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62723392))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63772032))))[name = string("encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63774144)))];
+            tensor<fp16, [1, 49, 1024]> linear_22_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_k_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_22_cast_fp16")];
+            tensor<int32, [4]> var_920 = const()[name = string("op_920"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_9_cast_fp16 = reshape(shape = var_920, x = linear_22_cast_fp16)[name = string("k_9_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63776256))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64824896))))[name = string("encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64827008)))];
+            tensor<fp16, [1, 49, 1024]> linear_23_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_v_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_23_cast_fp16")];
+            tensor<int32, [4]> var_925 = const()[name = string("op_925"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_5_cast_fp16 = reshape(shape = var_925, x = linear_23_cast_fp16)[name = string("v_5_cast_fp16")];
+            tensor<int32, [4]> value_13_perm_0 = const()[name = string("value_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64829120)))];
+            tensor<fp16, [1, 7, 8, 128]> var_938_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_u_to_fp16)[name = string("op_938_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64831232)))];
+            tensor<fp16, [1, 7, 8, 128]> var_940_cast_fp16 = add(x = q_13_cast_fp16, y = encoder_layers_2_self_attn_pos_bias_v_to_fp16)[name = string("op_940_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_5_perm_0 = const()[name = string("q_with_bias_v_5_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_59_transpose_x_0 = const()[name = string("x_59_transpose_x_0"), val = bool(false)];
+            bool x_59_transpose_y_0 = const()[name = string("x_59_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_942_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64833344))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64932736))))[name = string("op_942_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_940_cast_fp16)[name = string("transpose_344")];
+            tensor<fp16, [1, 8, 7, 97]> x_59_cast_fp16 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = op_942_to_fp16_quantized)[name = string("x_59_cast_fp16")];
+            tensor<int32, [8]> x_61_pad_0 = const()[name = string("x_61_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_61_mode_0 = const()[name = string("x_61_mode_0"), val = string("constant")];
+            fp16 const_105_to_fp16 = const()[name = string("const_105_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_61_cast_fp16 = pad(constant_val = const_105_to_fp16, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59_cast_fp16)[name = string("x_61_cast_fp16")];
+            tensor<int32, [4]> var_950 = const()[name = string("op_950"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_63_cast_fp16 = reshape(shape = var_950, x = x_61_cast_fp16)[name = string("x_63_cast_fp16")];
+            tensor<int32, [4]> var_954_begin_0 = const()[name = string("op_954_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_954_end_0 = const()[name = string("op_954_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_954_end_mask_0 = const()[name = string("op_954_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_954_cast_fp16 = slice_by_index(begin = var_954_begin_0, end = var_954_end_0, end_mask = var_954_end_mask_0, x = x_63_cast_fp16)[name = string("op_954_cast_fp16")];
+            tensor<int32, [4]> var_955 = const()[name = string("op_955"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_9_cast_fp16 = reshape(shape = var_955, x = var_954_cast_fp16)[name = string("matrix_bd_9_cast_fp16")];
+            bool matrix_ac_5_transpose_x_0 = const()[name = string("matrix_ac_5_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_5_transpose_y_0 = const()[name = string("matrix_ac_5_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_101 = transpose(perm = transpose_101_perm_0, x = k_9_cast_fp16)[name = string("transpose_342")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_100 = transpose(perm = transpose_100_perm_0, x = var_938_cast_fp16)[name = string("transpose_343")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_5_cast_fp16 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_100, y = transpose_101)[name = string("matrix_ac_5_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_11_begin_0 = const()[name = string("matrix_bd_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_11_end_0 = const()[name = string("matrix_bd_11_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_11_end_mask_0 = const()[name = string("matrix_bd_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_11_cast_fp16 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9_cast_fp16)[name = string("matrix_bd_11_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_964_cast_fp16 = add(x = matrix_ac_5_cast_fp16, y = matrix_bd_11_cast_fp16)[name = string("op_964_cast_fp16")];
+            fp16 _inversed_scores_9_y_0_to_fp16 = const()[name = string("_inversed_scores_9_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_9_cast_fp16 = mul(x = var_964_cast_fp16, y = _inversed_scores_9_y_0_to_fp16)[name = string("_inversed_scores_9_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_11_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_9_cast_fp16, cond = mask_11)[name = string("scores_11_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_970_cast_fp16 = softmax(axis = var_59, x = scores_11_cast_fp16)[name = string("op_970_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_145_cast_fp16 = select(a = var_44_to_fp16, b = var_970_cast_fp16, cond = mask_11)[name = string("input_145_cast_fp16")];
+            bool x_65_transpose_x_0 = const()[name = string("x_65_transpose_x_0"), val = bool(false)];
+            bool x_65_transpose_y_0 = const()[name = string("x_65_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_13_cast_fp16 = transpose(perm = value_13_perm_0, x = v_5_cast_fp16)[name = string("transpose_341")];
+            tensor<fp16, [1, 8, 7, 128]> x_65_cast_fp16 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145_cast_fp16, y = value_13_cast_fp16)[name = string("x_65_cast_fp16")];
+            tensor<int32, [4]> var_974_perm_0 = const()[name = string("op_974_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_975 = const()[name = string("op_975"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_974_cast_fp16 = transpose(perm = var_974_perm_0, x = x_65_cast_fp16)[name = string("transpose_340")];
+            tensor<fp16, [1, 7, 1024]> input_147_cast_fp16 = reshape(shape = var_975, x = var_974_cast_fp16)[name = string("input_147_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64933056))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65719552))))[name = string("encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_2_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_2_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65719744)))];
+            tensor<fp16, [1, 7, 1024]> linear_25_cast_fp16 = linear(bias = encoder_layers_2_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_2_self_attn_linear_out_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = string("linear_25_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_151_cast_fp16 = add(x = input_141_cast_fp16, y = linear_25_cast_fp16)[name = string("input_151_cast_fp16")];
+            tensor<int32, [1]> x_69_axes_0 = const()[name = string("x_69_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65721856)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65723968)))];
+            tensor<fp16, [1, 7, 1024]> x_69_cast_fp16 = layer_norm(axes = x_69_axes_0, beta = encoder_layers_2_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_conv_weight_to_fp16, x = input_151_cast_fp16)[name = string("x_69_cast_fp16")];
+            tensor<int32, [3]> input_153_perm_0 = const()[name = string("input_153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_155_pad_type_0 = const()[name = string("input_155_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_155_strides_0 = const()[name = string("input_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_155_pad_0 = const()[name = string("input_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_155_dilations_0 = const()[name = string("input_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_155_groups_0 = const()[name = string("input_155_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65726080))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67823296))))[name = string("encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_153_cast_fp16 = transpose(perm = input_153_perm_0, x = x_69_cast_fp16)[name = string("transpose_339")];
+            tensor<fp16, [1, 2048, 7]> input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = encoder_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")];
+            int32 x_71_split_num_splits_0 = const()[name = string("x_71_split_num_splits_0"), val = int32(2)];
+            int32 x_71_split_axis_0 = const()[name = string("x_71_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_71_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_71_split_cast_fp16_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155_cast_fp16)[name = string("x_71_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_71_split_1_sigmoid_cast_fp16 = sigmoid(x = x_71_split_cast_fp16_1)[name = string("x_71_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_71_cast_fp16 = mul(x = x_71_split_cast_fp16_0, y = x_71_split_1_sigmoid_cast_fp16)[name = string("x_71_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_157_cast_fp16 = select(a = var_44_to_fp16, b = x_71_cast_fp16, cond = var_575)[name = string("input_157_cast_fp16")];
+            bool new_x_11_interleave_0 = const()[name = string("new_x_11_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_11_cast_fp16 = concat(axis = var_59, interleave = new_x_11_interleave_0, values = (cache_11_cast_fp16, input_157_cast_fp16))[name = string("new_x_11_cast_fp16")];
+            tensor<int32, [3]> var_1014_begin_0 = const()[name = string("op_1014_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_1014_end_0 = const()[name = string("op_1014_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_1014_end_mask_0 = const()[name = string("op_1014_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1014_cast_fp16 = slice_by_index(begin = var_1014_begin_0, end = var_1014_end_0, end_mask = var_1014_end_mask_0, x = new_x_11_cast_fp16)[name = string("op_1014_cast_fp16")];
+            string x_73_pad_type_0 = const()[name = string("x_73_pad_type_0"), val = string("valid")];
+            int32 x_73_groups_0 = const()[name = string("x_73_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_73_strides_0 = const()[name = string("x_73_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_73_pad_0 = const()[name = string("x_73_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_73_dilations_0 = const()[name = string("x_73_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67827456))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67836736))))[name = string("encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_73_cast_fp16 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = encoder_layers_2_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_11_cast_fp16)[name = string("x_73_cast_fp16")];
+            tensor<int32, [3]> input_159_perm_0 = const()[name = string("input_159_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_75_axes_0 = const()[name = string("x_75_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67838848)))];
+            tensor<fp16, [1024]> encoder_layers_2_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_2_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67840960)))];
+            tensor<fp16, [1, 7, 1024]> input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_73_cast_fp16)[name = string("transpose_338")];
+            tensor<fp16, [1, 7, 1024]> x_75_cast_fp16 = layer_norm(axes = x_75_axes_0, beta = encoder_layers_2_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_conv_batch_norm_weight_to_fp16, x = input_159_cast_fp16)[name = string("x_75_cast_fp16")];
+            tensor<int32, [3]> input_161_perm_0 = const()[name = string("input_161_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_161_cast_fp16 = transpose(perm = input_161_perm_0, x = x_75_cast_fp16)[name = string("transpose_337")];
+            tensor<fp16, [1, 1024, 7]> input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = string("input_163_cast_fp16")];
+            string x_77_pad_type_0 = const()[name = string("x_77_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_77_strides_0 = const()[name = string("x_77_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_77_pad_0 = const()[name = string("x_77_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_77_dilations_0 = const()[name = string("x_77_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_77_groups_0 = const()[name = string("x_77_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67843072))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68891712))))[name = string("encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = encoder_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("x_77_cast_fp16")];
+            tensor<int32, [3]> input_165_perm_0 = const()[name = string("input_165_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_165_cast_fp16 = transpose(perm = input_165_perm_0, x = x_77_cast_fp16)[name = string("transpose_336")];
+            tensor<fp16, [1, 7, 1024]> input_167_cast_fp16 = add(x = input_151_cast_fp16, y = input_165_cast_fp16)[name = string("input_167_cast_fp16")];
+            tensor<int32, [1]> input_169_axes_0 = const()[name = string("input_169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68893824)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68895936)))];
+            tensor<fp16, [1, 7, 1024]> input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = encoder_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_feed_forward2_weight_to_fp16, x = input_167_cast_fp16)[name = string("input_169_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68898048))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72043840))))[name = string("encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_2_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72044032)))];
+            tensor<fp16, [1, 7, 4096]> linear_26_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear1_weight_to_fp16_palettized, x = input_169_cast_fp16)[name = string("linear_26_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_173_cast_fp16 = silu(x = linear_26_cast_fp16)[name = string("input_173_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72052288))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75198080))))[name = string("encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_2_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_2_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75198272)))];
+            tensor<fp16, [1, 7, 1024]> linear_27_cast_fp16 = linear(bias = encoder_layers_2_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_2_feed_forward2_linear2_weight_to_fp16_palettized, x = input_173_cast_fp16)[name = string("linear_27_cast_fp16")];
+            fp16 var_1057_to_fp16 = const()[name = string("op_1057_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1058_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_1057_to_fp16)[name = string("op_1058_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_179_cast_fp16 = add(x = input_167_cast_fp16, y = var_1058_cast_fp16)[name = string("input_179_cast_fp16")];
+            tensor<int32, [1]> input_181_axes_0 = const()[name = string("input_181_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_2_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_2_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75200384)))];
+            tensor<fp16, [1024]> encoder_layers_2_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_2_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75202496)))];
+            tensor<fp16, [1, 7, 1024]> input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = encoder_layers_2_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_2_norm_out_weight_to_fp16, x = input_179_cast_fp16)[name = string("input_181_cast_fp16")];
+            tensor<int32, [4]> cache_13_begin_0 = const()[name = string("cache_13_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_13_end_0 = const()[name = string("cache_13_end_0"), val = tensor<int32, [4]>([4, 1, 42, 1024])];
+            tensor<bool, [4]> cache_13_end_mask_0 = const()[name = string("cache_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_13_squeeze_mask_0 = const()[name = string("cache_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_13_cast_fp16 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_13_cast_fp16")];
+            tensor<int32, [4]> cache_15_begin_0 = const()[name = string("cache_15_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_15_end_0 = const()[name = string("cache_15_end_0"), val = tensor<int32, [4]>([4, 1, 1024, 8])];
+            tensor<bool, [4]> cache_15_end_mask_0 = const()[name = string("cache_15_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_15_squeeze_mask_0 = const()[name = string("cache_15_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_15_cast_fp16 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_15_cast_fp16")];
+            tensor<int32, [1]> input_183_axes_0 = const()[name = string("input_183_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75204608)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75206720)))];
+            tensor<fp16, [1, 7, 1024]> input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = encoder_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_feed_forward1_weight_to_fp16, x = input_181_cast_fp16)[name = string("input_183_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75208832))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78354624))))[name = string("encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_3_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78354816)))];
+            tensor<fp16, [1, 7, 4096]> linear_28_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear1_weight_to_fp16_palettized, x = input_183_cast_fp16)[name = string("linear_28_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_187_cast_fp16 = silu(x = linear_28_cast_fp16)[name = string("input_187_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78363072))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81508864))))[name = string("encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81509056)))];
+            tensor<fp16, [1, 7, 1024]> linear_29_cast_fp16 = linear(bias = encoder_layers_3_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward1_linear2_weight_to_fp16_palettized, x = input_187_cast_fp16)[name = string("linear_29_cast_fp16")];
+            fp16 var_1094_to_fp16 = const()[name = string("op_1094_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1095_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_1094_to_fp16)[name = string("op_1095_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_193_cast_fp16 = add(x = input_181_cast_fp16, y = var_1095_cast_fp16)[name = string("input_193_cast_fp16")];
+            tensor<int32, [1]> key_7_axes_0 = const()[name = string("key_7_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81511168)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81513280)))];
+            tensor<fp16, [1, 7, 1024]> key_7_cast_fp16 = layer_norm(axes = key_7_axes_0, beta = encoder_layers_3_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_self_att_weight_to_fp16, x = input_193_cast_fp16)[name = string("key_7_cast_fp16")];
+            bool input_195_interleave_0 = const()[name = string("input_195_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_195_cast_fp16 = concat(axis = var_68, interleave = input_195_interleave_0, values = (cache_13_cast_fp16, key_7_cast_fp16))[name = string("input_195_cast_fp16")];
+            tensor<int32, [3]> var_1117_begin_0 = const()[name = string("op_1117_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_1117_end_0 = const()[name = string("op_1117_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1117_end_mask_0 = const()[name = string("op_1117_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_1117_cast_fp16 = slice_by_index(begin = var_1117_begin_0, end = var_1117_end_0, end_mask = var_1117_end_mask_0, x = cache_13_cast_fp16)[name = string("op_1117_cast_fp16")];
+            bool var_1123_interleave_0 = const()[name = string("op_1123_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1123_cast_fp16 = concat(axis = var_68, interleave = var_1123_interleave_0, values = (var_1117_cast_fp16, key_7_cast_fp16))[name = string("op_1123_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81515392))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82301888))))[name = string("encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82302080)))];
+            tensor<fp16, [1, 7, 1024]> linear_30_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_q_weight_to_fp16_palettized, x = key_7_cast_fp16)[name = string("linear_30_cast_fp16")];
+            tensor<int32, [4]> var_1128 = const()[name = string("op_1128"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_19_cast_fp16 = reshape(shape = var_1128, x = linear_30_cast_fp16)[name = string("q_19_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82304192))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83090688))))[name = string("encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83090880)))];
+            tensor<fp16, [1, 49, 1024]> linear_31_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_k_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_31_cast_fp16")];
+            tensor<int32, [4]> var_1133 = const()[name = string("op_1133"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_13_cast_fp16 = reshape(shape = var_1133, x = linear_31_cast_fp16)[name = string("k_13_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83092992))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83879488))))[name = string("encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83879680)))];
+            tensor<fp16, [1, 49, 1024]> linear_32_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_v_weight_to_fp16_palettized, x = input_195_cast_fp16)[name = string("linear_32_cast_fp16")];
+            tensor<int32, [4]> var_1138 = const()[name = string("op_1138"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_7_cast_fp16 = reshape(shape = var_1138, x = linear_32_cast_fp16)[name = string("v_7_cast_fp16")];
+            tensor<int32, [4]> value_15_perm_0 = const()[name = string("value_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83881792)))];
+            tensor<fp16, [1, 7, 8, 128]> var_1151_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_u_to_fp16)[name = string("op_1151_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83883904)))];
+            tensor<fp16, [1, 7, 8, 128]> var_1153_cast_fp16 = add(x = q_19_cast_fp16, y = encoder_layers_3_self_attn_pos_bias_v_to_fp16)[name = string("op_1153_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_7_perm_0 = const()[name = string("q_with_bias_v_7_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_85_transpose_x_0 = const()[name = string("x_85_transpose_x_0"), val = bool(false)];
+            bool x_85_transpose_y_0 = const()[name = string("x_85_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_1155_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83886016))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83985408))))[name = string("op_1155_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_1153_cast_fp16)[name = string("transpose_335")];
+            tensor<fp16, [1, 8, 7, 97]> x_85_cast_fp16 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = op_1155_to_fp16_quantized)[name = string("x_85_cast_fp16")];
+            tensor<int32, [8]> x_87_pad_0 = const()[name = string("x_87_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_87_mode_0 = const()[name = string("x_87_mode_0"), val = string("constant")];
+            fp16 const_118_to_fp16 = const()[name = string("const_118_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_87_cast_fp16 = pad(constant_val = const_118_to_fp16, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85_cast_fp16)[name = string("x_87_cast_fp16")];
+            tensor<int32, [4]> var_1163 = const()[name = string("op_1163"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_89_cast_fp16 = reshape(shape = var_1163, x = x_87_cast_fp16)[name = string("x_89_cast_fp16")];
+            tensor<int32, [4]> var_1167_begin_0 = const()[name = string("op_1167_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1167_end_0 = const()[name = string("op_1167_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_1167_end_mask_0 = const()[name = string("op_1167_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_1167_cast_fp16 = slice_by_index(begin = var_1167_begin_0, end = var_1167_end_0, end_mask = var_1167_end_mask_0, x = x_89_cast_fp16)[name = string("op_1167_cast_fp16")];
+            tensor<int32, [4]> var_1168 = const()[name = string("op_1168"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_13_cast_fp16 = reshape(shape = var_1168, x = var_1167_cast_fp16)[name = string("matrix_bd_13_cast_fp16")];
+            bool matrix_ac_7_transpose_x_0 = const()[name = string("matrix_ac_7_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_7_transpose_y_0 = const()[name = string("matrix_ac_7_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_103 = transpose(perm = transpose_103_perm_0, x = k_13_cast_fp16)[name = string("transpose_333")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_102 = transpose(perm = transpose_102_perm_0, x = var_1151_cast_fp16)[name = string("transpose_334")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_7_cast_fp16 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_102, y = transpose_103)[name = string("matrix_ac_7_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_15_begin_0 = const()[name = string("matrix_bd_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_15_end_0 = const()[name = string("matrix_bd_15_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_15_end_mask_0 = const()[name = string("matrix_bd_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_15_cast_fp16 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13_cast_fp16)[name = string("matrix_bd_15_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_1177_cast_fp16 = add(x = matrix_ac_7_cast_fp16, y = matrix_bd_15_cast_fp16)[name = string("op_1177_cast_fp16")];
+            fp16 _inversed_scores_13_y_0_to_fp16 = const()[name = string("_inversed_scores_13_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_13_cast_fp16 = mul(x = var_1177_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = string("_inversed_scores_13_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_15_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_13_cast_fp16, cond = mask_11)[name = string("scores_15_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_1183_cast_fp16 = softmax(axis = var_59, x = scores_15_cast_fp16)[name = string("op_1183_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_197_cast_fp16 = select(a = var_44_to_fp16, b = var_1183_cast_fp16, cond = mask_11)[name = string("input_197_cast_fp16")];
+            bool x_91_transpose_x_0 = const()[name = string("x_91_transpose_x_0"), val = bool(false)];
+            bool x_91_transpose_y_0 = const()[name = string("x_91_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_15_cast_fp16 = transpose(perm = value_15_perm_0, x = v_7_cast_fp16)[name = string("transpose_332")];
+            tensor<fp16, [1, 8, 7, 128]> x_91_cast_fp16 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197_cast_fp16, y = value_15_cast_fp16)[name = string("x_91_cast_fp16")];
+            tensor<int32, [4]> var_1187_perm_0 = const()[name = string("op_1187_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1188 = const()[name = string("op_1188"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_1187_cast_fp16 = transpose(perm = var_1187_perm_0, x = x_91_cast_fp16)[name = string("transpose_331")];
+            tensor<fp16, [1, 7, 1024]> input_199_cast_fp16 = reshape(shape = var_1188, x = var_1187_cast_fp16)[name = string("input_199_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83985728))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84772224))))[name = string("encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_3_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84772416)))];
+            tensor<fp16, [1, 7, 1024]> linear_34_cast_fp16 = linear(bias = encoder_layers_3_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_3_self_attn_linear_out_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = string("linear_34_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_203_cast_fp16 = add(x = input_193_cast_fp16, y = linear_34_cast_fp16)[name = string("input_203_cast_fp16")];
+            tensor<int32, [1]> x_95_axes_0 = const()[name = string("x_95_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84774528)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84776640)))];
+            tensor<fp16, [1, 7, 1024]> x_95_cast_fp16 = layer_norm(axes = x_95_axes_0, beta = encoder_layers_3_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_conv_weight_to_fp16, x = input_203_cast_fp16)[name = string("x_95_cast_fp16")];
+            tensor<int32, [3]> input_205_perm_0 = const()[name = string("input_205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_207_pad_type_0 = const()[name = string("input_207_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_207_strides_0 = const()[name = string("input_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_207_pad_0 = const()[name = string("input_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_207_dilations_0 = const()[name = string("input_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_207_groups_0 = const()[name = string("input_207_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84778752))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86875968))))[name = string("encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_205_cast_fp16 = transpose(perm = input_205_perm_0, x = x_95_cast_fp16)[name = string("transpose_330")];
+            tensor<fp16, [1, 2048, 7]> input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = encoder_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("input_207_cast_fp16")];
+            int32 x_97_split_num_splits_0 = const()[name = string("x_97_split_num_splits_0"), val = int32(2)];
+            int32 x_97_split_axis_0 = const()[name = string("x_97_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_97_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_97_split_cast_fp16_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207_cast_fp16)[name = string("x_97_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_97_split_1_sigmoid_cast_fp16 = sigmoid(x = x_97_split_cast_fp16_1)[name = string("x_97_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_97_cast_fp16 = mul(x = x_97_split_cast_fp16_0, y = x_97_split_1_sigmoid_cast_fp16)[name = string("x_97_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_209_cast_fp16 = select(a = var_44_to_fp16, b = x_97_cast_fp16, cond = var_575)[name = string("input_209_cast_fp16")];
+            bool new_x_15_interleave_0 = const()[name = string("new_x_15_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_15_cast_fp16 = concat(axis = var_59, interleave = new_x_15_interleave_0, values = (cache_15_cast_fp16, input_209_cast_fp16))[name = string("new_x_15_cast_fp16")];
+            tensor<int32, [3]> var_1227_begin_0 = const()[name = string("op_1227_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_1227_end_0 = const()[name = string("op_1227_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_1227_end_mask_0 = const()[name = string("op_1227_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1227_cast_fp16 = slice_by_index(begin = var_1227_begin_0, end = var_1227_end_0, end_mask = var_1227_end_mask_0, x = new_x_15_cast_fp16)[name = string("op_1227_cast_fp16")];
+            string x_99_pad_type_0 = const()[name = string("x_99_pad_type_0"), val = string("valid")];
+            int32 x_99_groups_0 = const()[name = string("x_99_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_99_strides_0 = const()[name = string("x_99_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_99_pad_0 = const()[name = string("x_99_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_99_dilations_0 = const()[name = string("x_99_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86880128))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86889408))))[name = string("encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_99_cast_fp16 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = encoder_layers_3_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_15_cast_fp16)[name = string("x_99_cast_fp16")];
+            tensor<int32, [3]> input_211_perm_0 = const()[name = string("input_211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_101_axes_0 = const()[name = string("x_101_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86891520)))];
+            tensor<fp16, [1024]> encoder_layers_3_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_3_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86893632)))];
+            tensor<fp16, [1, 7, 1024]> input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_99_cast_fp16)[name = string("transpose_329")];
+            tensor<fp16, [1, 7, 1024]> x_101_cast_fp16 = layer_norm(axes = x_101_axes_0, beta = encoder_layers_3_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_conv_batch_norm_weight_to_fp16, x = input_211_cast_fp16)[name = string("x_101_cast_fp16")];
+            tensor<int32, [3]> input_213_perm_0 = const()[name = string("input_213_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_213_cast_fp16 = transpose(perm = input_213_perm_0, x = x_101_cast_fp16)[name = string("transpose_328")];
+            tensor<fp16, [1, 1024, 7]> input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = string("input_215_cast_fp16")];
+            string x_103_pad_type_0 = const()[name = string("x_103_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_103_strides_0 = const()[name = string("x_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_103_pad_0 = const()[name = string("x_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_103_dilations_0 = const()[name = string("x_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_103_groups_0 = const()[name = string("x_103_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86895744))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87944384))))[name = string("encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_103_cast_fp16 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = encoder_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("x_103_cast_fp16")];
+            tensor<int32, [3]> input_217_perm_0 = const()[name = string("input_217_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_217_cast_fp16 = transpose(perm = input_217_perm_0, x = x_103_cast_fp16)[name = string("transpose_327")];
+            tensor<fp16, [1, 7, 1024]> input_219_cast_fp16 = add(x = input_203_cast_fp16, y = input_217_cast_fp16)[name = string("input_219_cast_fp16")];
+            tensor<int32, [1]> input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87946496)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87948608)))];
+            tensor<fp16, [1, 7, 1024]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_feed_forward2_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87950720))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91096512))))[name = string("encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_3_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91096704)))];
+            tensor<fp16, [1, 7, 4096]> linear_35_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear1_weight_to_fp16_palettized, x = input_221_cast_fp16)[name = string("linear_35_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_225_cast_fp16 = silu(x = linear_35_cast_fp16)[name = string("input_225_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(91104960))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94250752))))[name = string("encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_3_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_3_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94250944)))];
+            tensor<fp16, [1, 7, 1024]> linear_36_cast_fp16 = linear(bias = encoder_layers_3_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_3_feed_forward2_linear2_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = string("linear_36_cast_fp16")];
+            fp16 var_1270_to_fp16 = const()[name = string("op_1270_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1271_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_1270_to_fp16)[name = string("op_1271_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_231_cast_fp16 = add(x = input_219_cast_fp16, y = var_1271_cast_fp16)[name = string("input_231_cast_fp16")];
+            tensor<int32, [1]> input_233_axes_0 = const()[name = string("input_233_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_3_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_3_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94253056)))];
+            tensor<fp16, [1024]> encoder_layers_3_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_3_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94255168)))];
+            tensor<fp16, [1, 7, 1024]> input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = encoder_layers_3_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_3_norm_out_weight_to_fp16, x = input_231_cast_fp16)[name = string("input_233_cast_fp16")];
+            tensor<int32, [4]> cache_17_begin_0 = const()[name = string("cache_17_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_17_end_0 = const()[name = string("cache_17_end_0"), val = tensor<int32, [4]>([5, 1, 42, 1024])];
+            tensor<bool, [4]> cache_17_end_mask_0 = const()[name = string("cache_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_17_squeeze_mask_0 = const()[name = string("cache_17_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_17_cast_fp16 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_17_cast_fp16")];
+            tensor<int32, [4]> cache_19_begin_0 = const()[name = string("cache_19_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_19_end_0 = const()[name = string("cache_19_end_0"), val = tensor<int32, [4]>([5, 1, 1024, 8])];
+            tensor<bool, [4]> cache_19_end_mask_0 = const()[name = string("cache_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_19_squeeze_mask_0 = const()[name = string("cache_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_19_cast_fp16 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_19_cast_fp16")];
+            tensor<int32, [1]> input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94257280)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94259392)))];
+            tensor<fp16, [1, 7, 1024]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_feed_forward1_weight_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94261504))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97407296))))[name = string("encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_4_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97407488)))];
+            tensor<fp16, [1, 7, 4096]> linear_37_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear1_weight_to_fp16_palettized, x = input_235_cast_fp16)[name = string("linear_37_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_239_cast_fp16 = silu(x = linear_37_cast_fp16)[name = string("input_239_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97415744))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100561536))))[name = string("encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100561728)))];
+            tensor<fp16, [1, 7, 1024]> linear_38_cast_fp16 = linear(bias = encoder_layers_4_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward1_linear2_weight_to_fp16_palettized, x = input_239_cast_fp16)[name = string("linear_38_cast_fp16")];
+            fp16 var_1307_to_fp16 = const()[name = string("op_1307_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1308_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_1307_to_fp16)[name = string("op_1308_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_245_cast_fp16 = add(x = input_233_cast_fp16, y = var_1308_cast_fp16)[name = string("input_245_cast_fp16")];
+            tensor<int32, [1]> key_9_axes_0 = const()[name = string("key_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100563840)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100565952)))];
+            tensor<fp16, [1, 7, 1024]> key_9_cast_fp16 = layer_norm(axes = key_9_axes_0, beta = encoder_layers_4_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_self_att_weight_to_fp16, x = input_245_cast_fp16)[name = string("key_9_cast_fp16")];
+            bool input_247_interleave_0 = const()[name = string("input_247_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_247_cast_fp16 = concat(axis = var_68, interleave = input_247_interleave_0, values = (cache_17_cast_fp16, key_9_cast_fp16))[name = string("input_247_cast_fp16")];
+            tensor<int32, [3]> var_1330_begin_0 = const()[name = string("op_1330_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_1330_end_0 = const()[name = string("op_1330_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1330_end_mask_0 = const()[name = string("op_1330_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_1330_cast_fp16 = slice_by_index(begin = var_1330_begin_0, end = var_1330_end_0, end_mask = var_1330_end_mask_0, x = cache_17_cast_fp16)[name = string("op_1330_cast_fp16")];
+            bool var_1336_interleave_0 = const()[name = string("op_1336_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1336_cast_fp16 = concat(axis = var_68, interleave = var_1336_interleave_0, values = (var_1330_cast_fp16, key_9_cast_fp16))[name = string("op_1336_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100568064))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101354560))))[name = string("encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101354752)))];
+            tensor<fp16, [1, 7, 1024]> linear_39_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_q_weight_to_fp16_palettized, x = key_9_cast_fp16)[name = string("linear_39_cast_fp16")];
+            tensor<int32, [4]> var_1341 = const()[name = string("op_1341"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_25_cast_fp16 = reshape(shape = var_1341, x = linear_39_cast_fp16)[name = string("q_25_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101356864))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102143360))))[name = string("encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102143552)))];
+            tensor<fp16, [1, 49, 1024]> linear_40_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_k_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_40_cast_fp16")];
+            tensor<int32, [4]> var_1346 = const()[name = string("op_1346"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_17_cast_fp16 = reshape(shape = var_1346, x = linear_40_cast_fp16)[name = string("k_17_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102145664))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102932160))))[name = string("encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102932352)))];
+            tensor<fp16, [1, 49, 1024]> linear_41_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_v_weight_to_fp16_palettized, x = input_247_cast_fp16)[name = string("linear_41_cast_fp16")];
+            tensor<int32, [4]> var_1351 = const()[name = string("op_1351"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_9_cast_fp16 = reshape(shape = var_1351, x = linear_41_cast_fp16)[name = string("v_9_cast_fp16")];
+            tensor<int32, [4]> value_17_perm_0 = const()[name = string("value_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102934464)))];
+            tensor<fp16, [1, 7, 8, 128]> var_1364_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_u_to_fp16)[name = string("op_1364_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102936576)))];
+            tensor<fp16, [1, 7, 8, 128]> var_1366_cast_fp16 = add(x = q_25_cast_fp16, y = encoder_layers_4_self_attn_pos_bias_v_to_fp16)[name = string("op_1366_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_9_perm_0 = const()[name = string("q_with_bias_v_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_111_transpose_x_0 = const()[name = string("x_111_transpose_x_0"), val = bool(false)];
+            bool x_111_transpose_y_0 = const()[name = string("x_111_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_1368_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102938688))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103038080))))[name = string("op_1368_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1366_cast_fp16)[name = string("transpose_326")];
+            tensor<fp16, [1, 8, 7, 97]> x_111_cast_fp16 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = op_1368_to_fp16_quantized)[name = string("x_111_cast_fp16")];
+            tensor<int32, [8]> x_113_pad_0 = const()[name = string("x_113_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_113_mode_0 = const()[name = string("x_113_mode_0"), val = string("constant")];
+            fp16 const_131_to_fp16 = const()[name = string("const_131_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_113_cast_fp16 = pad(constant_val = const_131_to_fp16, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111_cast_fp16)[name = string("x_113_cast_fp16")];
+            tensor<int32, [4]> var_1376 = const()[name = string("op_1376"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_115_cast_fp16 = reshape(shape = var_1376, x = x_113_cast_fp16)[name = string("x_115_cast_fp16")];
+            tensor<int32, [4]> var_1380_begin_0 = const()[name = string("op_1380_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1380_end_0 = const()[name = string("op_1380_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_1380_end_mask_0 = const()[name = string("op_1380_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_1380_cast_fp16 = slice_by_index(begin = var_1380_begin_0, end = var_1380_end_0, end_mask = var_1380_end_mask_0, x = x_115_cast_fp16)[name = string("op_1380_cast_fp16")];
+            tensor<int32, [4]> var_1381 = const()[name = string("op_1381"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_17_cast_fp16 = reshape(shape = var_1381, x = var_1380_cast_fp16)[name = string("matrix_bd_17_cast_fp16")];
+            bool matrix_ac_9_transpose_x_0 = const()[name = string("matrix_ac_9_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_9_transpose_y_0 = const()[name = string("matrix_ac_9_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_105 = transpose(perm = transpose_105_perm_0, x = k_17_cast_fp16)[name = string("transpose_324")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1364_cast_fp16)[name = string("transpose_325")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_9_cast_fp16 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_104, y = transpose_105)[name = string("matrix_ac_9_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_19_begin_0 = const()[name = string("matrix_bd_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_19_end_0 = const()[name = string("matrix_bd_19_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_19_end_mask_0 = const()[name = string("matrix_bd_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_19_cast_fp16 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17_cast_fp16)[name = string("matrix_bd_19_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_1390_cast_fp16 = add(x = matrix_ac_9_cast_fp16, y = matrix_bd_19_cast_fp16)[name = string("op_1390_cast_fp16")];
+            fp16 _inversed_scores_17_y_0_to_fp16 = const()[name = string("_inversed_scores_17_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_17_cast_fp16 = mul(x = var_1390_cast_fp16, y = _inversed_scores_17_y_0_to_fp16)[name = string("_inversed_scores_17_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_19_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_17_cast_fp16, cond = mask_11)[name = string("scores_19_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_1396_cast_fp16 = softmax(axis = var_59, x = scores_19_cast_fp16)[name = string("op_1396_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_249_cast_fp16 = select(a = var_44_to_fp16, b = var_1396_cast_fp16, cond = mask_11)[name = string("input_249_cast_fp16")];
+            bool x_117_transpose_x_0 = const()[name = string("x_117_transpose_x_0"), val = bool(false)];
+            bool x_117_transpose_y_0 = const()[name = string("x_117_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_17_cast_fp16 = transpose(perm = value_17_perm_0, x = v_9_cast_fp16)[name = string("transpose_323")];
+            tensor<fp16, [1, 8, 7, 128]> x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249_cast_fp16, y = value_17_cast_fp16)[name = string("x_117_cast_fp16")];
+            tensor<int32, [4]> var_1400_perm_0 = const()[name = string("op_1400_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1401 = const()[name = string("op_1401"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_1400_cast_fp16 = transpose(perm = var_1400_perm_0, x = x_117_cast_fp16)[name = string("transpose_322")];
+            tensor<fp16, [1, 7, 1024]> input_251_cast_fp16 = reshape(shape = var_1401, x = var_1400_cast_fp16)[name = string("input_251_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103038400))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103824896))))[name = string("encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_4_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103825088)))];
+            tensor<fp16, [1, 7, 1024]> linear_43_cast_fp16 = linear(bias = encoder_layers_4_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_4_self_attn_linear_out_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = string("linear_43_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_255_cast_fp16 = add(x = input_245_cast_fp16, y = linear_43_cast_fp16)[name = string("input_255_cast_fp16")];
+            tensor<int32, [1]> x_121_axes_0 = const()[name = string("x_121_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103827200)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103829312)))];
+            tensor<fp16, [1, 7, 1024]> x_121_cast_fp16 = layer_norm(axes = x_121_axes_0, beta = encoder_layers_4_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_conv_weight_to_fp16, x = input_255_cast_fp16)[name = string("x_121_cast_fp16")];
+            tensor<int32, [3]> input_257_perm_0 = const()[name = string("input_257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_259_pad_type_0 = const()[name = string("input_259_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_259_strides_0 = const()[name = string("input_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_259_pad_0 = const()[name = string("input_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_259_dilations_0 = const()[name = string("input_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_259_groups_0 = const()[name = string("input_259_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(103831424))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105928640))))[name = string("encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_257_cast_fp16 = transpose(perm = input_257_perm_0, x = x_121_cast_fp16)[name = string("transpose_321")];
+            tensor<fp16, [1, 2048, 7]> input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = encoder_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_257_cast_fp16)[name = string("input_259_cast_fp16")];
+            int32 x_123_split_num_splits_0 = const()[name = string("x_123_split_num_splits_0"), val = int32(2)];
+            int32 x_123_split_axis_0 = const()[name = string("x_123_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_123_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_123_split_cast_fp16_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259_cast_fp16)[name = string("x_123_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_123_split_1_sigmoid_cast_fp16 = sigmoid(x = x_123_split_cast_fp16_1)[name = string("x_123_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_123_cast_fp16 = mul(x = x_123_split_cast_fp16_0, y = x_123_split_1_sigmoid_cast_fp16)[name = string("x_123_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_261_cast_fp16 = select(a = var_44_to_fp16, b = x_123_cast_fp16, cond = var_575)[name = string("input_261_cast_fp16")];
+            bool new_x_19_interleave_0 = const()[name = string("new_x_19_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_19_cast_fp16 = concat(axis = var_59, interleave = new_x_19_interleave_0, values = (cache_19_cast_fp16, input_261_cast_fp16))[name = string("new_x_19_cast_fp16")];
+            tensor<int32, [3]> var_1440_begin_0 = const()[name = string("op_1440_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_1440_end_0 = const()[name = string("op_1440_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_1440_end_mask_0 = const()[name = string("op_1440_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1440_cast_fp16 = slice_by_index(begin = var_1440_begin_0, end = var_1440_end_0, end_mask = var_1440_end_mask_0, x = new_x_19_cast_fp16)[name = string("op_1440_cast_fp16")];
+            string x_125_pad_type_0 = const()[name = string("x_125_pad_type_0"), val = string("valid")];
+            int32 x_125_groups_0 = const()[name = string("x_125_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_125_strides_0 = const()[name = string("x_125_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_125_pad_0 = const()[name = string("x_125_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_125_dilations_0 = const()[name = string("x_125_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105932800))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105942080))))[name = string("encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = encoder_layers_4_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_19_cast_fp16)[name = string("x_125_cast_fp16")];
+            tensor<int32, [3]> input_263_perm_0 = const()[name = string("input_263_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_127_axes_0 = const()[name = string("x_127_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105944192)))];
+            tensor<fp16, [1024]> encoder_layers_4_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_4_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105946304)))];
+            tensor<fp16, [1, 7, 1024]> input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_125_cast_fp16)[name = string("transpose_320")];
+            tensor<fp16, [1, 7, 1024]> x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = encoder_layers_4_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_conv_batch_norm_weight_to_fp16, x = input_263_cast_fp16)[name = string("x_127_cast_fp16")];
+            tensor<int32, [3]> input_265_perm_0 = const()[name = string("input_265_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_265_cast_fp16 = transpose(perm = input_265_perm_0, x = x_127_cast_fp16)[name = string("transpose_319")];
+            tensor<fp16, [1, 1024, 7]> input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = string("input_267_cast_fp16")];
+            string x_129_pad_type_0 = const()[name = string("x_129_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_129_strides_0 = const()[name = string("x_129_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_129_pad_0 = const()[name = string("x_129_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_129_dilations_0 = const()[name = string("x_129_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_129_groups_0 = const()[name = string("x_129_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105948416))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106997056))))[name = string("encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_129_cast_fp16 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = encoder_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = string("x_129_cast_fp16")];
+            tensor<int32, [3]> input_269_perm_0 = const()[name = string("input_269_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_269_cast_fp16 = transpose(perm = input_269_perm_0, x = x_129_cast_fp16)[name = string("transpose_318")];
+            tensor<fp16, [1, 7, 1024]> input_271_cast_fp16 = add(x = input_255_cast_fp16, y = input_269_cast_fp16)[name = string("input_271_cast_fp16")];
+            tensor<int32, [1]> input_273_axes_0 = const()[name = string("input_273_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106999168)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107001280)))];
+            tensor<fp16, [1, 7, 1024]> input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = encoder_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_feed_forward2_weight_to_fp16, x = input_271_cast_fp16)[name = string("input_273_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107003392))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110149184))))[name = string("encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_4_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110149376)))];
+            tensor<fp16, [1, 7, 4096]> linear_44_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear1_weight_to_fp16_palettized, x = input_273_cast_fp16)[name = string("linear_44_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_277_cast_fp16 = silu(x = linear_44_cast_fp16)[name = string("input_277_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110157632))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113303424))))[name = string("encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_4_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_4_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113303616)))];
+            tensor<fp16, [1, 7, 1024]> linear_45_cast_fp16 = linear(bias = encoder_layers_4_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_4_feed_forward2_linear2_weight_to_fp16_palettized, x = input_277_cast_fp16)[name = string("linear_45_cast_fp16")];
+            fp16 var_1483_to_fp16 = const()[name = string("op_1483_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1484_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1483_to_fp16)[name = string("op_1484_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_283_cast_fp16 = add(x = input_271_cast_fp16, y = var_1484_cast_fp16)[name = string("input_283_cast_fp16")];
+            tensor<int32, [1]> input_285_axes_0 = const()[name = string("input_285_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_4_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_4_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113305728)))];
+            tensor<fp16, [1024]> encoder_layers_4_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_4_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113307840)))];
+            tensor<fp16, [1, 7, 1024]> input_285_cast_fp16 = layer_norm(axes = input_285_axes_0, beta = encoder_layers_4_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_4_norm_out_weight_to_fp16, x = input_283_cast_fp16)[name = string("input_285_cast_fp16")];
+            tensor<int32, [4]> cache_21_begin_0 = const()[name = string("cache_21_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_21_end_0 = const()[name = string("cache_21_end_0"), val = tensor<int32, [4]>([6, 1, 42, 1024])];
+            tensor<bool, [4]> cache_21_end_mask_0 = const()[name = string("cache_21_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_21_squeeze_mask_0 = const()[name = string("cache_21_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_21_cast_fp16 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_21_cast_fp16")];
+            tensor<int32, [4]> cache_23_begin_0 = const()[name = string("cache_23_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_23_end_0 = const()[name = string("cache_23_end_0"), val = tensor<int32, [4]>([6, 1, 1024, 8])];
+            tensor<bool, [4]> cache_23_end_mask_0 = const()[name = string("cache_23_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_23_squeeze_mask_0 = const()[name = string("cache_23_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_23_cast_fp16 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_23_cast_fp16")];
+            tensor<int32, [1]> input_287_axes_0 = const()[name = string("input_287_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113309952)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113312064)))];
+            tensor<fp16, [1, 7, 1024]> input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = encoder_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_feed_forward1_weight_to_fp16, x = input_285_cast_fp16)[name = string("input_287_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(113314176))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116459968))))[name = string("encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_5_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116460160)))];
+            tensor<fp16, [1, 7, 4096]> linear_46_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear1_weight_to_fp16_palettized, x = input_287_cast_fp16)[name = string("linear_46_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_291_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("input_291_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116468416))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119614208))))[name = string("encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119614400)))];
+            tensor<fp16, [1, 7, 1024]> linear_47_cast_fp16 = linear(bias = encoder_layers_5_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward1_linear2_weight_to_fp16_palettized, x = input_291_cast_fp16)[name = string("linear_47_cast_fp16")];
+            fp16 var_1520_to_fp16 = const()[name = string("op_1520_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1521_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1520_to_fp16)[name = string("op_1521_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_297_cast_fp16 = add(x = input_285_cast_fp16, y = var_1521_cast_fp16)[name = string("input_297_cast_fp16")];
+            tensor<int32, [1]> key_11_axes_0 = const()[name = string("key_11_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119616512)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119618624)))];
+            tensor<fp16, [1, 7, 1024]> key_11_cast_fp16 = layer_norm(axes = key_11_axes_0, beta = encoder_layers_5_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_self_att_weight_to_fp16, x = input_297_cast_fp16)[name = string("key_11_cast_fp16")];
+            bool input_299_interleave_0 = const()[name = string("input_299_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_299_cast_fp16 = concat(axis = var_68, interleave = input_299_interleave_0, values = (cache_21_cast_fp16, key_11_cast_fp16))[name = string("input_299_cast_fp16")];
+            tensor<int32, [3]> var_1543_begin_0 = const()[name = string("op_1543_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_1543_end_0 = const()[name = string("op_1543_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1543_end_mask_0 = const()[name = string("op_1543_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_1543_cast_fp16 = slice_by_index(begin = var_1543_begin_0, end = var_1543_end_0, end_mask = var_1543_end_mask_0, x = cache_21_cast_fp16)[name = string("op_1543_cast_fp16")];
+            bool var_1549_interleave_0 = const()[name = string("op_1549_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1549_cast_fp16 = concat(axis = var_68, interleave = var_1549_interleave_0, values = (var_1543_cast_fp16, key_11_cast_fp16))[name = string("op_1549_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119620736))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120407232))))[name = string("encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120407424)))];
+            tensor<fp16, [1, 7, 1024]> linear_48_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_q_weight_to_fp16_palettized, x = key_11_cast_fp16)[name = string("linear_48_cast_fp16")];
+            tensor<int32, [4]> var_1554 = const()[name = string("op_1554"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_31_cast_fp16 = reshape(shape = var_1554, x = linear_48_cast_fp16)[name = string("q_31_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120409536))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121196032))))[name = string("encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121196224)))];
+            tensor<fp16, [1, 49, 1024]> linear_49_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_k_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_49_cast_fp16")];
+            tensor<int32, [4]> var_1559 = const()[name = string("op_1559"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_21_cast_fp16 = reshape(shape = var_1559, x = linear_49_cast_fp16)[name = string("k_21_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121198336))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121984832))))[name = string("encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121985024)))];
+            tensor<fp16, [1, 49, 1024]> linear_50_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_v_weight_to_fp16_palettized, x = input_299_cast_fp16)[name = string("linear_50_cast_fp16")];
+            tensor<int32, [4]> var_1564 = const()[name = string("op_1564"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_11_cast_fp16 = reshape(shape = var_1564, x = linear_50_cast_fp16)[name = string("v_11_cast_fp16")];
+            tensor<int32, [4]> value_19_perm_0 = const()[name = string("value_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121987136)))];
+            tensor<fp16, [1, 7, 8, 128]> var_1577_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_u_to_fp16)[name = string("op_1577_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121989248)))];
+            tensor<fp16, [1, 7, 8, 128]> var_1579_cast_fp16 = add(x = q_31_cast_fp16, y = encoder_layers_5_self_attn_pos_bias_v_to_fp16)[name = string("op_1579_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_11_perm_0 = const()[name = string("q_with_bias_v_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_137_transpose_x_0 = const()[name = string("x_137_transpose_x_0"), val = bool(false)];
+            bool x_137_transpose_y_0 = const()[name = string("x_137_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_1581_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121991360))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122090752))))[name = string("op_1581_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1579_cast_fp16)[name = string("transpose_317")];
+            tensor<fp16, [1, 8, 7, 97]> x_137_cast_fp16 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = op_1581_to_fp16_quantized)[name = string("x_137_cast_fp16")];
+            tensor<int32, [8]> x_139_pad_0 = const()[name = string("x_139_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_139_mode_0 = const()[name = string("x_139_mode_0"), val = string("constant")];
+            fp16 const_144_to_fp16 = const()[name = string("const_144_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_139_cast_fp16 = pad(constant_val = const_144_to_fp16, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137_cast_fp16)[name = string("x_139_cast_fp16")];
+            tensor<int32, [4]> var_1589 = const()[name = string("op_1589"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_141_cast_fp16 = reshape(shape = var_1589, x = x_139_cast_fp16)[name = string("x_141_cast_fp16")];
+            tensor<int32, [4]> var_1593_begin_0 = const()[name = string("op_1593_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1593_end_0 = const()[name = string("op_1593_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_1593_end_mask_0 = const()[name = string("op_1593_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_1593_cast_fp16 = slice_by_index(begin = var_1593_begin_0, end = var_1593_end_0, end_mask = var_1593_end_mask_0, x = x_141_cast_fp16)[name = string("op_1593_cast_fp16")];
+            tensor<int32, [4]> var_1594 = const()[name = string("op_1594"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_21_cast_fp16 = reshape(shape = var_1594, x = var_1593_cast_fp16)[name = string("matrix_bd_21_cast_fp16")];
+            bool matrix_ac_11_transpose_x_0 = const()[name = string("matrix_ac_11_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_11_transpose_y_0 = const()[name = string("matrix_ac_11_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_107 = transpose(perm = transpose_107_perm_0, x = k_21_cast_fp16)[name = string("transpose_315")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_106 = transpose(perm = transpose_106_perm_0, x = var_1577_cast_fp16)[name = string("transpose_316")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_11_cast_fp16 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_106, y = transpose_107)[name = string("matrix_ac_11_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_23_begin_0 = const()[name = string("matrix_bd_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_23_end_0 = const()[name = string("matrix_bd_23_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_23_end_mask_0 = const()[name = string("matrix_bd_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_23_cast_fp16 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21_cast_fp16)[name = string("matrix_bd_23_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_1603_cast_fp16 = add(x = matrix_ac_11_cast_fp16, y = matrix_bd_23_cast_fp16)[name = string("op_1603_cast_fp16")];
+            fp16 _inversed_scores_21_y_0_to_fp16 = const()[name = string("_inversed_scores_21_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_21_cast_fp16 = mul(x = var_1603_cast_fp16, y = _inversed_scores_21_y_0_to_fp16)[name = string("_inversed_scores_21_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_23_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_21_cast_fp16, cond = mask_11)[name = string("scores_23_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_1609_cast_fp16 = softmax(axis = var_59, x = scores_23_cast_fp16)[name = string("op_1609_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_301_cast_fp16 = select(a = var_44_to_fp16, b = var_1609_cast_fp16, cond = mask_11)[name = string("input_301_cast_fp16")];
+            bool x_143_transpose_x_0 = const()[name = string("x_143_transpose_x_0"), val = bool(false)];
+            bool x_143_transpose_y_0 = const()[name = string("x_143_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_19_cast_fp16 = transpose(perm = value_19_perm_0, x = v_11_cast_fp16)[name = string("transpose_314")];
+            tensor<fp16, [1, 8, 7, 128]> x_143_cast_fp16 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301_cast_fp16, y = value_19_cast_fp16)[name = string("x_143_cast_fp16")];
+            tensor<int32, [4]> var_1613_perm_0 = const()[name = string("op_1613_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1614 = const()[name = string("op_1614"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_1613_cast_fp16 = transpose(perm = var_1613_perm_0, x = x_143_cast_fp16)[name = string("transpose_313")];
+            tensor<fp16, [1, 7, 1024]> input_303_cast_fp16 = reshape(shape = var_1614, x = var_1613_cast_fp16)[name = string("input_303_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122091072))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122877568))))[name = string("encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_5_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122877760)))];
+            tensor<fp16, [1, 7, 1024]> linear_52_cast_fp16 = linear(bias = encoder_layers_5_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_5_self_attn_linear_out_weight_to_fp16_palettized, x = input_303_cast_fp16)[name = string("linear_52_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_307_cast_fp16 = add(x = input_297_cast_fp16, y = linear_52_cast_fp16)[name = string("input_307_cast_fp16")];
+            tensor<int32, [1]> x_147_axes_0 = const()[name = string("x_147_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122879872)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122881984)))];
+            tensor<fp16, [1, 7, 1024]> x_147_cast_fp16 = layer_norm(axes = x_147_axes_0, beta = encoder_layers_5_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_conv_weight_to_fp16, x = input_307_cast_fp16)[name = string("x_147_cast_fp16")];
+            tensor<int32, [3]> input_309_perm_0 = const()[name = string("input_309_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_311_pad_type_0 = const()[name = string("input_311_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_311_strides_0 = const()[name = string("input_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_311_pad_0 = const()[name = string("input_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_311_dilations_0 = const()[name = string("input_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_311_groups_0 = const()[name = string("input_311_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122884096))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124981312))))[name = string("encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_309_cast_fp16 = transpose(perm = input_309_perm_0, x = x_147_cast_fp16)[name = string("transpose_312")];
+            tensor<fp16, [1, 2048, 7]> input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = encoder_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_309_cast_fp16)[name = string("input_311_cast_fp16")];
+            int32 x_149_split_num_splits_0 = const()[name = string("x_149_split_num_splits_0"), val = int32(2)];
+            int32 x_149_split_axis_0 = const()[name = string("x_149_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_149_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_149_split_cast_fp16_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311_cast_fp16)[name = string("x_149_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_149_split_1_sigmoid_cast_fp16 = sigmoid(x = x_149_split_cast_fp16_1)[name = string("x_149_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_149_cast_fp16 = mul(x = x_149_split_cast_fp16_0, y = x_149_split_1_sigmoid_cast_fp16)[name = string("x_149_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_313_cast_fp16 = select(a = var_44_to_fp16, b = x_149_cast_fp16, cond = var_575)[name = string("input_313_cast_fp16")];
+            bool new_x_23_interleave_0 = const()[name = string("new_x_23_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_23_cast_fp16 = concat(axis = var_59, interleave = new_x_23_interleave_0, values = (cache_23_cast_fp16, input_313_cast_fp16))[name = string("new_x_23_cast_fp16")];
+            tensor<int32, [3]> var_1653_begin_0 = const()[name = string("op_1653_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_1653_end_0 = const()[name = string("op_1653_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_1653_end_mask_0 = const()[name = string("op_1653_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1653_cast_fp16 = slice_by_index(begin = var_1653_begin_0, end = var_1653_end_0, end_mask = var_1653_end_mask_0, x = new_x_23_cast_fp16)[name = string("op_1653_cast_fp16")];
+            string x_151_pad_type_0 = const()[name = string("x_151_pad_type_0"), val = string("valid")];
+            int32 x_151_groups_0 = const()[name = string("x_151_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_151_strides_0 = const()[name = string("x_151_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_151_pad_0 = const()[name = string("x_151_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_151_dilations_0 = const()[name = string("x_151_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124985472))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124994752))))[name = string("encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_151_cast_fp16 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = encoder_layers_5_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_23_cast_fp16)[name = string("x_151_cast_fp16")];
+            tensor<int32, [3]> input_315_perm_0 = const()[name = string("input_315_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_153_axes_0 = const()[name = string("x_153_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124996864)))];
+            tensor<fp16, [1024]> encoder_layers_5_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_5_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124998976)))];
+            tensor<fp16, [1, 7, 1024]> input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_151_cast_fp16)[name = string("transpose_311")];
+            tensor<fp16, [1, 7, 1024]> x_153_cast_fp16 = layer_norm(axes = x_153_axes_0, beta = encoder_layers_5_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_conv_batch_norm_weight_to_fp16, x = input_315_cast_fp16)[name = string("x_153_cast_fp16")];
+            tensor<int32, [3]> input_317_perm_0 = const()[name = string("input_317_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_317_cast_fp16 = transpose(perm = input_317_perm_0, x = x_153_cast_fp16)[name = string("transpose_310")];
+            tensor<fp16, [1, 1024, 7]> input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = string("input_319_cast_fp16")];
+            string x_155_pad_type_0 = const()[name = string("x_155_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_155_strides_0 = const()[name = string("x_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_155_pad_0 = const()[name = string("x_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_155_dilations_0 = const()[name = string("x_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_155_groups_0 = const()[name = string("x_155_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125001088))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126049728))))[name = string("encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_155_cast_fp16 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = encoder_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_319_cast_fp16)[name = string("x_155_cast_fp16")];
+            tensor<int32, [3]> input_321_perm_0 = const()[name = string("input_321_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_321_cast_fp16 = transpose(perm = input_321_perm_0, x = x_155_cast_fp16)[name = string("transpose_309")];
+            tensor<fp16, [1, 7, 1024]> input_323_cast_fp16 = add(x = input_307_cast_fp16, y = input_321_cast_fp16)[name = string("input_323_cast_fp16")];
+            tensor<int32, [1]> input_325_axes_0 = const()[name = string("input_325_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126051840)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126053952)))];
+            tensor<fp16, [1, 7, 1024]> input_325_cast_fp16 = layer_norm(axes = input_325_axes_0, beta = encoder_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_feed_forward2_weight_to_fp16, x = input_323_cast_fp16)[name = string("input_325_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126056064))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129201856))))[name = string("encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_5_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129202048)))];
+            tensor<fp16, [1, 7, 4096]> linear_53_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear1_weight_to_fp16_palettized, x = input_325_cast_fp16)[name = string("linear_53_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_329_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("input_329_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(129210304))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132356096))))[name = string("encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_5_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_5_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132356288)))];
+            tensor<fp16, [1, 7, 1024]> linear_54_cast_fp16 = linear(bias = encoder_layers_5_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_5_feed_forward2_linear2_weight_to_fp16_palettized, x = input_329_cast_fp16)[name = string("linear_54_cast_fp16")];
+            fp16 var_1696_to_fp16 = const()[name = string("op_1696_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1697_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1696_to_fp16)[name = string("op_1697_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_335_cast_fp16 = add(x = input_323_cast_fp16, y = var_1697_cast_fp16)[name = string("input_335_cast_fp16")];
+            tensor<int32, [1]> input_337_axes_0 = const()[name = string("input_337_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_5_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_5_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132358400)))];
+            tensor<fp16, [1024]> encoder_layers_5_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_5_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132360512)))];
+            tensor<fp16, [1, 7, 1024]> input_337_cast_fp16 = layer_norm(axes = input_337_axes_0, beta = encoder_layers_5_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_5_norm_out_weight_to_fp16, x = input_335_cast_fp16)[name = string("input_337_cast_fp16")];
+            tensor<int32, [4]> cache_25_begin_0 = const()[name = string("cache_25_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_25_end_0 = const()[name = string("cache_25_end_0"), val = tensor<int32, [4]>([7, 1, 42, 1024])];
+            tensor<bool, [4]> cache_25_end_mask_0 = const()[name = string("cache_25_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_25_squeeze_mask_0 = const()[name = string("cache_25_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_25_cast_fp16 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_25_cast_fp16")];
+            tensor<int32, [4]> cache_27_begin_0 = const()[name = string("cache_27_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_27_end_0 = const()[name = string("cache_27_end_0"), val = tensor<int32, [4]>([7, 1, 1024, 8])];
+            tensor<bool, [4]> cache_27_end_mask_0 = const()[name = string("cache_27_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_27_squeeze_mask_0 = const()[name = string("cache_27_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_27_cast_fp16 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_27_cast_fp16")];
+            tensor<int32, [1]> input_339_axes_0 = const()[name = string("input_339_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132362624)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132364736)))];
+            tensor<fp16, [1, 7, 1024]> input_339_cast_fp16 = layer_norm(axes = input_339_axes_0, beta = encoder_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_feed_forward1_weight_to_fp16, x = input_337_cast_fp16)[name = string("input_339_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132366848))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135512640))))[name = string("encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_6_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135512832)))];
+            tensor<fp16, [1, 7, 4096]> linear_55_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear1_weight_to_fp16_palettized, x = input_339_cast_fp16)[name = string("linear_55_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_343_cast_fp16 = silu(x = linear_55_cast_fp16)[name = string("input_343_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135521088))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138666880))))[name = string("encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138667072)))];
+            tensor<fp16, [1, 7, 1024]> linear_56_cast_fp16 = linear(bias = encoder_layers_6_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward1_linear2_weight_to_fp16_palettized, x = input_343_cast_fp16)[name = string("linear_56_cast_fp16")];
+            fp16 var_1733_to_fp16 = const()[name = string("op_1733_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1734_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1733_to_fp16)[name = string("op_1734_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_349_cast_fp16 = add(x = input_337_cast_fp16, y = var_1734_cast_fp16)[name = string("input_349_cast_fp16")];
+            tensor<int32, [1]> key_13_axes_0 = const()[name = string("key_13_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138669184)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138671296)))];
+            tensor<fp16, [1, 7, 1024]> key_13_cast_fp16 = layer_norm(axes = key_13_axes_0, beta = encoder_layers_6_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_self_att_weight_to_fp16, x = input_349_cast_fp16)[name = string("key_13_cast_fp16")];
+            bool input_351_interleave_0 = const()[name = string("input_351_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_351_cast_fp16 = concat(axis = var_68, interleave = input_351_interleave_0, values = (cache_25_cast_fp16, key_13_cast_fp16))[name = string("input_351_cast_fp16")];
+            tensor<int32, [3]> var_1756_begin_0 = const()[name = string("op_1756_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_1756_end_0 = const()[name = string("op_1756_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1756_end_mask_0 = const()[name = string("op_1756_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_1756_cast_fp16 = slice_by_index(begin = var_1756_begin_0, end = var_1756_end_0, end_mask = var_1756_end_mask_0, x = cache_25_cast_fp16)[name = string("op_1756_cast_fp16")];
+            bool var_1762_interleave_0 = const()[name = string("op_1762_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1762_cast_fp16 = concat(axis = var_68, interleave = var_1762_interleave_0, values = (var_1756_cast_fp16, key_13_cast_fp16))[name = string("op_1762_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(138673408))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139459904))))[name = string("encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139460096)))];
+            tensor<fp16, [1, 7, 1024]> linear_57_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_q_weight_to_fp16_palettized, x = key_13_cast_fp16)[name = string("linear_57_cast_fp16")];
+            tensor<int32, [4]> var_1767 = const()[name = string("op_1767"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_37_cast_fp16 = reshape(shape = var_1767, x = linear_57_cast_fp16)[name = string("q_37_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(139462208))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140248704))))[name = string("encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140248896)))];
+            tensor<fp16, [1, 49, 1024]> linear_58_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_k_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_58_cast_fp16")];
+            tensor<int32, [4]> var_1772 = const()[name = string("op_1772"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_25_cast_fp16 = reshape(shape = var_1772, x = linear_58_cast_fp16)[name = string("k_25_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140251008))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141037504))))[name = string("encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141037696)))];
+            tensor<fp16, [1, 49, 1024]> linear_59_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_v_weight_to_fp16_palettized, x = input_351_cast_fp16)[name = string("linear_59_cast_fp16")];
+            tensor<int32, [4]> var_1777 = const()[name = string("op_1777"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_13_cast_fp16 = reshape(shape = var_1777, x = linear_59_cast_fp16)[name = string("v_13_cast_fp16")];
+            tensor<int32, [4]> value_21_perm_0 = const()[name = string("value_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141039808)))];
+            tensor<fp16, [1, 7, 8, 128]> var_1790_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_u_to_fp16)[name = string("op_1790_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141041920)))];
+            tensor<fp16, [1, 7, 8, 128]> var_1792_cast_fp16 = add(x = q_37_cast_fp16, y = encoder_layers_6_self_attn_pos_bias_v_to_fp16)[name = string("op_1792_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_13_perm_0 = const()[name = string("q_with_bias_v_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_163_transpose_x_0 = const()[name = string("x_163_transpose_x_0"), val = bool(false)];
+            bool x_163_transpose_y_0 = const()[name = string("x_163_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_1794_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141044032))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141143424))))[name = string("op_1794_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1792_cast_fp16)[name = string("transpose_308")];
+            tensor<fp16, [1, 8, 7, 97]> x_163_cast_fp16 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = op_1794_to_fp16_quantized)[name = string("x_163_cast_fp16")];
+            tensor<int32, [8]> x_165_pad_0 = const()[name = string("x_165_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_165_mode_0 = const()[name = string("x_165_mode_0"), val = string("constant")];
+            fp16 const_157_to_fp16 = const()[name = string("const_157_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_165_cast_fp16 = pad(constant_val = const_157_to_fp16, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163_cast_fp16)[name = string("x_165_cast_fp16")];
+            tensor<int32, [4]> var_1802 = const()[name = string("op_1802"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_167_cast_fp16 = reshape(shape = var_1802, x = x_165_cast_fp16)[name = string("x_167_cast_fp16")];
+            tensor<int32, [4]> var_1806_begin_0 = const()[name = string("op_1806_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1806_end_0 = const()[name = string("op_1806_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_1806_end_mask_0 = const()[name = string("op_1806_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_1806_cast_fp16 = slice_by_index(begin = var_1806_begin_0, end = var_1806_end_0, end_mask = var_1806_end_mask_0, x = x_167_cast_fp16)[name = string("op_1806_cast_fp16")];
+            tensor<int32, [4]> var_1807 = const()[name = string("op_1807"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_25_cast_fp16 = reshape(shape = var_1807, x = var_1806_cast_fp16)[name = string("matrix_bd_25_cast_fp16")];
+            bool matrix_ac_13_transpose_x_0 = const()[name = string("matrix_ac_13_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_13_transpose_y_0 = const()[name = string("matrix_ac_13_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_109 = transpose(perm = transpose_109_perm_0, x = k_25_cast_fp16)[name = string("transpose_306")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1790_cast_fp16)[name = string("transpose_307")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_13_cast_fp16 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_108, y = transpose_109)[name = string("matrix_ac_13_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_27_begin_0 = const()[name = string("matrix_bd_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_27_end_0 = const()[name = string("matrix_bd_27_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_27_end_mask_0 = const()[name = string("matrix_bd_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_27_cast_fp16 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25_cast_fp16)[name = string("matrix_bd_27_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_1816_cast_fp16 = add(x = matrix_ac_13_cast_fp16, y = matrix_bd_27_cast_fp16)[name = string("op_1816_cast_fp16")];
+            fp16 _inversed_scores_25_y_0_to_fp16 = const()[name = string("_inversed_scores_25_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_25_cast_fp16 = mul(x = var_1816_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = string("_inversed_scores_25_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_27_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_25_cast_fp16, cond = mask_11)[name = string("scores_27_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_1822_cast_fp16 = softmax(axis = var_59, x = scores_27_cast_fp16)[name = string("op_1822_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_353_cast_fp16 = select(a = var_44_to_fp16, b = var_1822_cast_fp16, cond = mask_11)[name = string("input_353_cast_fp16")];
+            bool x_169_transpose_x_0 = const()[name = string("x_169_transpose_x_0"), val = bool(false)];
+            bool x_169_transpose_y_0 = const()[name = string("x_169_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_21_cast_fp16 = transpose(perm = value_21_perm_0, x = v_13_cast_fp16)[name = string("transpose_305")];
+            tensor<fp16, [1, 8, 7, 128]> x_169_cast_fp16 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353_cast_fp16, y = value_21_cast_fp16)[name = string("x_169_cast_fp16")];
+            tensor<int32, [4]> var_1826_perm_0 = const()[name = string("op_1826_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1827 = const()[name = string("op_1827"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_1826_cast_fp16 = transpose(perm = var_1826_perm_0, x = x_169_cast_fp16)[name = string("transpose_304")];
+            tensor<fp16, [1, 7, 1024]> input_355_cast_fp16 = reshape(shape = var_1827, x = var_1826_cast_fp16)[name = string("input_355_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141143744))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141930240))))[name = string("encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_6_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141930432)))];
+            tensor<fp16, [1, 7, 1024]> linear_61_cast_fp16 = linear(bias = encoder_layers_6_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_6_self_attn_linear_out_weight_to_fp16_palettized, x = input_355_cast_fp16)[name = string("linear_61_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_359_cast_fp16 = add(x = input_349_cast_fp16, y = linear_61_cast_fp16)[name = string("input_359_cast_fp16")];
+            tensor<int32, [1]> x_173_axes_0 = const()[name = string("x_173_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141932544)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141934656)))];
+            tensor<fp16, [1, 7, 1024]> x_173_cast_fp16 = layer_norm(axes = x_173_axes_0, beta = encoder_layers_6_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_conv_weight_to_fp16, x = input_359_cast_fp16)[name = string("x_173_cast_fp16")];
+            tensor<int32, [3]> input_361_perm_0 = const()[name = string("input_361_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_363_pad_type_0 = const()[name = string("input_363_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_363_strides_0 = const()[name = string("input_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_363_pad_0 = const()[name = string("input_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_363_dilations_0 = const()[name = string("input_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_363_groups_0 = const()[name = string("input_363_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(141936768))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144033984))))[name = string("encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_361_cast_fp16 = transpose(perm = input_361_perm_0, x = x_173_cast_fp16)[name = string("transpose_303")];
+            tensor<fp16, [1, 2048, 7]> input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = encoder_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_361_cast_fp16)[name = string("input_363_cast_fp16")];
+            int32 x_175_split_num_splits_0 = const()[name = string("x_175_split_num_splits_0"), val = int32(2)];
+            int32 x_175_split_axis_0 = const()[name = string("x_175_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_175_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_175_split_cast_fp16_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363_cast_fp16)[name = string("x_175_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_175_split_1_sigmoid_cast_fp16 = sigmoid(x = x_175_split_cast_fp16_1)[name = string("x_175_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_175_cast_fp16 = mul(x = x_175_split_cast_fp16_0, y = x_175_split_1_sigmoid_cast_fp16)[name = string("x_175_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_365_cast_fp16 = select(a = var_44_to_fp16, b = x_175_cast_fp16, cond = var_575)[name = string("input_365_cast_fp16")];
+            bool new_x_27_interleave_0 = const()[name = string("new_x_27_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_27_cast_fp16 = concat(axis = var_59, interleave = new_x_27_interleave_0, values = (cache_27_cast_fp16, input_365_cast_fp16))[name = string("new_x_27_cast_fp16")];
+            tensor<int32, [3]> var_1866_begin_0 = const()[name = string("op_1866_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_1866_end_0 = const()[name = string("op_1866_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_1866_end_mask_0 = const()[name = string("op_1866_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_1866_cast_fp16 = slice_by_index(begin = var_1866_begin_0, end = var_1866_end_0, end_mask = var_1866_end_mask_0, x = new_x_27_cast_fp16)[name = string("op_1866_cast_fp16")];
+            string x_177_pad_type_0 = const()[name = string("x_177_pad_type_0"), val = string("valid")];
+            int32 x_177_groups_0 = const()[name = string("x_177_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_177_strides_0 = const()[name = string("x_177_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_177_pad_0 = const()[name = string("x_177_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_177_dilations_0 = const()[name = string("x_177_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144038144))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144047424))))[name = string("encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_177_cast_fp16 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = encoder_layers_6_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_27_cast_fp16)[name = string("x_177_cast_fp16")];
+            tensor<int32, [3]> input_367_perm_0 = const()[name = string("input_367_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_179_axes_0 = const()[name = string("x_179_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144049536)))];
+            tensor<fp16, [1024]> encoder_layers_6_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_6_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144051648)))];
+            tensor<fp16, [1, 7, 1024]> input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_177_cast_fp16)[name = string("transpose_302")];
+            tensor<fp16, [1, 7, 1024]> x_179_cast_fp16 = layer_norm(axes = x_179_axes_0, beta = encoder_layers_6_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_conv_batch_norm_weight_to_fp16, x = input_367_cast_fp16)[name = string("x_179_cast_fp16")];
+            tensor<int32, [3]> input_369_perm_0 = const()[name = string("input_369_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_369_cast_fp16 = transpose(perm = input_369_perm_0, x = x_179_cast_fp16)[name = string("transpose_301")];
+            tensor<fp16, [1, 1024, 7]> input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = string("input_371_cast_fp16")];
+            string x_181_pad_type_0 = const()[name = string("x_181_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_181_strides_0 = const()[name = string("x_181_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_181_pad_0 = const()[name = string("x_181_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_181_dilations_0 = const()[name = string("x_181_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_181_groups_0 = const()[name = string("x_181_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144053760))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145102400))))[name = string("encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_181_cast_fp16 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = encoder_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_371_cast_fp16)[name = string("x_181_cast_fp16")];
+            tensor<int32, [3]> input_373_perm_0 = const()[name = string("input_373_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_373_cast_fp16 = transpose(perm = input_373_perm_0, x = x_181_cast_fp16)[name = string("transpose_300")];
+            tensor<fp16, [1, 7, 1024]> input_375_cast_fp16 = add(x = input_359_cast_fp16, y = input_373_cast_fp16)[name = string("input_375_cast_fp16")];
+            tensor<int32, [1]> input_377_axes_0 = const()[name = string("input_377_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145104512)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145106624)))];
+            tensor<fp16, [1, 7, 1024]> input_377_cast_fp16 = layer_norm(axes = input_377_axes_0, beta = encoder_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_feed_forward2_weight_to_fp16, x = input_375_cast_fp16)[name = string("input_377_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145108736))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148254528))))[name = string("encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_6_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148254720)))];
+            tensor<fp16, [1, 7, 4096]> linear_62_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear1_weight_to_fp16_palettized, x = input_377_cast_fp16)[name = string("linear_62_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_381_cast_fp16 = silu(x = linear_62_cast_fp16)[name = string("input_381_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148262976))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151408768))))[name = string("encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_6_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_6_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151408960)))];
+            tensor<fp16, [1, 7, 1024]> linear_63_cast_fp16 = linear(bias = encoder_layers_6_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_6_feed_forward2_linear2_weight_to_fp16_palettized, x = input_381_cast_fp16)[name = string("linear_63_cast_fp16")];
+            fp16 var_1909_to_fp16 = const()[name = string("op_1909_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1910_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1909_to_fp16)[name = string("op_1910_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_387_cast_fp16 = add(x = input_375_cast_fp16, y = var_1910_cast_fp16)[name = string("input_387_cast_fp16")];
+            tensor<int32, [1]> input_389_axes_0 = const()[name = string("input_389_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_6_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_6_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151411072)))];
+            tensor<fp16, [1024]> encoder_layers_6_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_6_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151413184)))];
+            tensor<fp16, [1, 7, 1024]> input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_layers_6_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_6_norm_out_weight_to_fp16, x = input_387_cast_fp16)[name = string("input_389_cast_fp16")];
+            tensor<int32, [4]> cache_29_begin_0 = const()[name = string("cache_29_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_29_end_0 = const()[name = string("cache_29_end_0"), val = tensor<int32, [4]>([8, 1, 42, 1024])];
+            tensor<bool, [4]> cache_29_end_mask_0 = const()[name = string("cache_29_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_29_squeeze_mask_0 = const()[name = string("cache_29_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_29_cast_fp16 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_29_cast_fp16")];
+            tensor<int32, [4]> cache_31_begin_0 = const()[name = string("cache_31_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_31_end_0 = const()[name = string("cache_31_end_0"), val = tensor<int32, [4]>([8, 1, 1024, 8])];
+            tensor<bool, [4]> cache_31_end_mask_0 = const()[name = string("cache_31_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_31_squeeze_mask_0 = const()[name = string("cache_31_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_31_cast_fp16 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_31_cast_fp16")];
+            tensor<int32, [1]> input_391_axes_0 = const()[name = string("input_391_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151415296)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151417408)))];
+            tensor<fp16, [1, 7, 1024]> input_391_cast_fp16 = layer_norm(axes = input_391_axes_0, beta = encoder_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_feed_forward1_weight_to_fp16, x = input_389_cast_fp16)[name = string("input_391_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151419520))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154565312))))[name = string("encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_7_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154565504)))];
+            tensor<fp16, [1, 7, 4096]> linear_64_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear1_weight_to_fp16_palettized, x = input_391_cast_fp16)[name = string("linear_64_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_395_cast_fp16 = silu(x = linear_64_cast_fp16)[name = string("input_395_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(154573760))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157719552))))[name = string("encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157719744)))];
+            tensor<fp16, [1, 7, 1024]> linear_65_cast_fp16 = linear(bias = encoder_layers_7_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward1_linear2_weight_to_fp16_palettized, x = input_395_cast_fp16)[name = string("linear_65_cast_fp16")];
+            fp16 var_1946_to_fp16 = const()[name = string("op_1946_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_1947_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1946_to_fp16)[name = string("op_1947_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_401_cast_fp16 = add(x = input_389_cast_fp16, y = var_1947_cast_fp16)[name = string("input_401_cast_fp16")];
+            tensor<int32, [1]> key_15_axes_0 = const()[name = string("key_15_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157721856)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157723968)))];
+            tensor<fp16, [1, 7, 1024]> key_15_cast_fp16 = layer_norm(axes = key_15_axes_0, beta = encoder_layers_7_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_self_att_weight_to_fp16, x = input_401_cast_fp16)[name = string("key_15_cast_fp16")];
+            bool input_403_interleave_0 = const()[name = string("input_403_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_403_cast_fp16 = concat(axis = var_68, interleave = input_403_interleave_0, values = (cache_29_cast_fp16, key_15_cast_fp16))[name = string("input_403_cast_fp16")];
+            tensor<int32, [3]> var_1969_begin_0 = const()[name = string("op_1969_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_1969_end_0 = const()[name = string("op_1969_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_1969_end_mask_0 = const()[name = string("op_1969_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_1969_cast_fp16 = slice_by_index(begin = var_1969_begin_0, end = var_1969_end_0, end_mask = var_1969_end_mask_0, x = cache_29_cast_fp16)[name = string("op_1969_cast_fp16")];
+            bool var_1975_interleave_0 = const()[name = string("op_1975_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_1975_cast_fp16 = concat(axis = var_68, interleave = var_1975_interleave_0, values = (var_1969_cast_fp16, key_15_cast_fp16))[name = string("op_1975_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157726080))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158512576))))[name = string("encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158512768)))];
+            tensor<fp16, [1, 7, 1024]> linear_66_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_q_weight_to_fp16_palettized, x = key_15_cast_fp16)[name = string("linear_66_cast_fp16")];
+            tensor<int32, [4]> var_1980 = const()[name = string("op_1980"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_43_cast_fp16 = reshape(shape = var_1980, x = linear_66_cast_fp16)[name = string("q_43_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(158514880))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159301376))))[name = string("encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159301568)))];
+            tensor<fp16, [1, 49, 1024]> linear_67_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_k_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_67_cast_fp16")];
+            tensor<int32, [4]> var_1985 = const()[name = string("op_1985"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_29_cast_fp16 = reshape(shape = var_1985, x = linear_67_cast_fp16)[name = string("k_29_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159303680))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160090176))))[name = string("encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160090368)))];
+            tensor<fp16, [1, 49, 1024]> linear_68_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_v_weight_to_fp16_palettized, x = input_403_cast_fp16)[name = string("linear_68_cast_fp16")];
+            tensor<int32, [4]> var_1990 = const()[name = string("op_1990"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_15_cast_fp16 = reshape(shape = var_1990, x = linear_68_cast_fp16)[name = string("v_15_cast_fp16")];
+            tensor<int32, [4]> value_23_perm_0 = const()[name = string("value_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160092480)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2003_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_u_to_fp16)[name = string("op_2003_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160094592)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2005_cast_fp16 = add(x = q_43_cast_fp16, y = encoder_layers_7_self_attn_pos_bias_v_to_fp16)[name = string("op_2005_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_15_perm_0 = const()[name = string("q_with_bias_v_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_189_transpose_x_0 = const()[name = string("x_189_transpose_x_0"), val = bool(false)];
+            bool x_189_transpose_y_0 = const()[name = string("x_189_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_2007_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160096704))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160196096))))[name = string("op_2007_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_2005_cast_fp16)[name = string("transpose_299")];
+            tensor<fp16, [1, 8, 7, 97]> x_189_cast_fp16 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = op_2007_to_fp16_quantized)[name = string("x_189_cast_fp16")];
+            tensor<int32, [8]> x_191_pad_0 = const()[name = string("x_191_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_191_mode_0 = const()[name = string("x_191_mode_0"), val = string("constant")];
+            fp16 const_170_to_fp16 = const()[name = string("const_170_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_191_cast_fp16 = pad(constant_val = const_170_to_fp16, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189_cast_fp16)[name = string("x_191_cast_fp16")];
+            tensor<int32, [4]> var_2015 = const()[name = string("op_2015"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_193_cast_fp16 = reshape(shape = var_2015, x = x_191_cast_fp16)[name = string("x_193_cast_fp16")];
+            tensor<int32, [4]> var_2019_begin_0 = const()[name = string("op_2019_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2019_end_0 = const()[name = string("op_2019_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_2019_end_mask_0 = const()[name = string("op_2019_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_2019_cast_fp16 = slice_by_index(begin = var_2019_begin_0, end = var_2019_end_0, end_mask = var_2019_end_mask_0, x = x_193_cast_fp16)[name = string("op_2019_cast_fp16")];
+            tensor<int32, [4]> var_2020 = const()[name = string("op_2020"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_29_cast_fp16 = reshape(shape = var_2020, x = var_2019_cast_fp16)[name = string("matrix_bd_29_cast_fp16")];
+            bool matrix_ac_15_transpose_x_0 = const()[name = string("matrix_ac_15_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_15_transpose_y_0 = const()[name = string("matrix_ac_15_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_111 = transpose(perm = transpose_111_perm_0, x = k_29_cast_fp16)[name = string("transpose_297")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_110 = transpose(perm = transpose_110_perm_0, x = var_2003_cast_fp16)[name = string("transpose_298")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_15_cast_fp16 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_110, y = transpose_111)[name = string("matrix_ac_15_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_31_begin_0 = const()[name = string("matrix_bd_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_31_end_0 = const()[name = string("matrix_bd_31_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_31_end_mask_0 = const()[name = string("matrix_bd_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_31_cast_fp16 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29_cast_fp16)[name = string("matrix_bd_31_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2029_cast_fp16 = add(x = matrix_ac_15_cast_fp16, y = matrix_bd_31_cast_fp16)[name = string("op_2029_cast_fp16")];
+            fp16 _inversed_scores_29_y_0_to_fp16 = const()[name = string("_inversed_scores_29_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_29_cast_fp16 = mul(x = var_2029_cast_fp16, y = _inversed_scores_29_y_0_to_fp16)[name = string("_inversed_scores_29_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_31_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_29_cast_fp16, cond = mask_11)[name = string("scores_31_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2035_cast_fp16 = softmax(axis = var_59, x = scores_31_cast_fp16)[name = string("op_2035_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_405_cast_fp16 = select(a = var_44_to_fp16, b = var_2035_cast_fp16, cond = mask_11)[name = string("input_405_cast_fp16")];
+            bool x_195_transpose_x_0 = const()[name = string("x_195_transpose_x_0"), val = bool(false)];
+            bool x_195_transpose_y_0 = const()[name = string("x_195_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_23_cast_fp16 = transpose(perm = value_23_perm_0, x = v_15_cast_fp16)[name = string("transpose_296")];
+            tensor<fp16, [1, 8, 7, 128]> x_195_cast_fp16 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405_cast_fp16, y = value_23_cast_fp16)[name = string("x_195_cast_fp16")];
+            tensor<int32, [4]> var_2039_perm_0 = const()[name = string("op_2039_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2040 = const()[name = string("op_2040"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_2039_cast_fp16 = transpose(perm = var_2039_perm_0, x = x_195_cast_fp16)[name = string("transpose_295")];
+            tensor<fp16, [1, 7, 1024]> input_407_cast_fp16 = reshape(shape = var_2040, x = var_2039_cast_fp16)[name = string("input_407_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160196416))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160982912))))[name = string("encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_7_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160983104)))];
+            tensor<fp16, [1, 7, 1024]> linear_70_cast_fp16 = linear(bias = encoder_layers_7_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_7_self_attn_linear_out_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = string("linear_70_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_411_cast_fp16 = add(x = input_401_cast_fp16, y = linear_70_cast_fp16)[name = string("input_411_cast_fp16")];
+            tensor<int32, [1]> x_199_axes_0 = const()[name = string("x_199_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160985216)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160987328)))];
+            tensor<fp16, [1, 7, 1024]> x_199_cast_fp16 = layer_norm(axes = x_199_axes_0, beta = encoder_layers_7_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_conv_weight_to_fp16, x = input_411_cast_fp16)[name = string("x_199_cast_fp16")];
+            tensor<int32, [3]> input_413_perm_0 = const()[name = string("input_413_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_415_pad_type_0 = const()[name = string("input_415_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_415_strides_0 = const()[name = string("input_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_415_pad_0 = const()[name = string("input_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_415_dilations_0 = const()[name = string("input_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_415_groups_0 = const()[name = string("input_415_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160989440))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163086656))))[name = string("encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_413_cast_fp16 = transpose(perm = input_413_perm_0, x = x_199_cast_fp16)[name = string("transpose_294")];
+            tensor<fp16, [1, 2048, 7]> input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = encoder_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_413_cast_fp16)[name = string("input_415_cast_fp16")];
+            int32 x_201_split_num_splits_0 = const()[name = string("x_201_split_num_splits_0"), val = int32(2)];
+            int32 x_201_split_axis_0 = const()[name = string("x_201_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_201_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_201_split_cast_fp16_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415_cast_fp16)[name = string("x_201_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_201_split_1_sigmoid_cast_fp16 = sigmoid(x = x_201_split_cast_fp16_1)[name = string("x_201_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_201_cast_fp16 = mul(x = x_201_split_cast_fp16_0, y = x_201_split_1_sigmoid_cast_fp16)[name = string("x_201_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_417_cast_fp16 = select(a = var_44_to_fp16, b = x_201_cast_fp16, cond = var_575)[name = string("input_417_cast_fp16")];
+            bool new_x_31_interleave_0 = const()[name = string("new_x_31_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_31_cast_fp16 = concat(axis = var_59, interleave = new_x_31_interleave_0, values = (cache_31_cast_fp16, input_417_cast_fp16))[name = string("new_x_31_cast_fp16")];
+            tensor<int32, [3]> var_2079_begin_0 = const()[name = string("op_2079_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_2079_end_0 = const()[name = string("op_2079_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_2079_end_mask_0 = const()[name = string("op_2079_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2079_cast_fp16 = slice_by_index(begin = var_2079_begin_0, end = var_2079_end_0, end_mask = var_2079_end_mask_0, x = new_x_31_cast_fp16)[name = string("op_2079_cast_fp16")];
+            string x_203_pad_type_0 = const()[name = string("x_203_pad_type_0"), val = string("valid")];
+            int32 x_203_groups_0 = const()[name = string("x_203_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_203_strides_0 = const()[name = string("x_203_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_203_pad_0 = const()[name = string("x_203_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_203_dilations_0 = const()[name = string("x_203_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163090816))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163100096))))[name = string("encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_203_cast_fp16 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = encoder_layers_7_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_31_cast_fp16)[name = string("x_203_cast_fp16")];
+            tensor<int32, [3]> input_419_perm_0 = const()[name = string("input_419_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_205_axes_0 = const()[name = string("x_205_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163102208)))];
+            tensor<fp16, [1024]> encoder_layers_7_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_7_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163104320)))];
+            tensor<fp16, [1, 7, 1024]> input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_203_cast_fp16)[name = string("transpose_293")];
+            tensor<fp16, [1, 7, 1024]> x_205_cast_fp16 = layer_norm(axes = x_205_axes_0, beta = encoder_layers_7_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_conv_batch_norm_weight_to_fp16, x = input_419_cast_fp16)[name = string("x_205_cast_fp16")];
+            tensor<int32, [3]> input_421_perm_0 = const()[name = string("input_421_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_421_cast_fp16 = transpose(perm = input_421_perm_0, x = x_205_cast_fp16)[name = string("transpose_292")];
+            tensor<fp16, [1, 1024, 7]> input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = string("input_423_cast_fp16")];
+            string x_207_pad_type_0 = const()[name = string("x_207_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_207_strides_0 = const()[name = string("x_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_207_pad_0 = const()[name = string("x_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_207_dilations_0 = const()[name = string("x_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_207_groups_0 = const()[name = string("x_207_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163106432))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164155072))))[name = string("encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_207_cast_fp16 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = encoder_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_423_cast_fp16)[name = string("x_207_cast_fp16")];
+            tensor<int32, [3]> input_425_perm_0 = const()[name = string("input_425_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_425_cast_fp16 = transpose(perm = input_425_perm_0, x = x_207_cast_fp16)[name = string("transpose_291")];
+            tensor<fp16, [1, 7, 1024]> input_427_cast_fp16 = add(x = input_411_cast_fp16, y = input_425_cast_fp16)[name = string("input_427_cast_fp16")];
+            tensor<int32, [1]> input_429_axes_0 = const()[name = string("input_429_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164157184)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164159296)))];
+            tensor<fp16, [1, 7, 1024]> input_429_cast_fp16 = layer_norm(axes = input_429_axes_0, beta = encoder_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_feed_forward2_weight_to_fp16, x = input_427_cast_fp16)[name = string("input_429_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(164161408))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167307200))))[name = string("encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_7_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167307392)))];
+            tensor<fp16, [1, 7, 4096]> linear_71_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear1_weight_to_fp16_palettized, x = input_429_cast_fp16)[name = string("linear_71_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_433_cast_fp16 = silu(x = linear_71_cast_fp16)[name = string("input_433_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167315648))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170461440))))[name = string("encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_7_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_7_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170461632)))];
+            tensor<fp16, [1, 7, 1024]> linear_72_cast_fp16 = linear(bias = encoder_layers_7_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_7_feed_forward2_linear2_weight_to_fp16_palettized, x = input_433_cast_fp16)[name = string("linear_72_cast_fp16")];
+            fp16 var_2122_to_fp16 = const()[name = string("op_2122_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_2123_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_2122_to_fp16)[name = string("op_2123_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_439_cast_fp16 = add(x = input_427_cast_fp16, y = var_2123_cast_fp16)[name = string("input_439_cast_fp16")];
+            tensor<int32, [1]> input_441_axes_0 = const()[name = string("input_441_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_7_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_7_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170463744)))];
+            tensor<fp16, [1024]> encoder_layers_7_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_7_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170465856)))];
+            tensor<fp16, [1, 7, 1024]> input_441_cast_fp16 = layer_norm(axes = input_441_axes_0, beta = encoder_layers_7_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_7_norm_out_weight_to_fp16, x = input_439_cast_fp16)[name = string("input_441_cast_fp16")];
+            tensor<int32, [4]> cache_33_begin_0 = const()[name = string("cache_33_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_33_end_0 = const()[name = string("cache_33_end_0"), val = tensor<int32, [4]>([9, 1, 42, 1024])];
+            tensor<bool, [4]> cache_33_end_mask_0 = const()[name = string("cache_33_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_33_squeeze_mask_0 = const()[name = string("cache_33_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_33_cast_fp16 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_33_cast_fp16")];
+            tensor<int32, [4]> cache_35_begin_0 = const()[name = string("cache_35_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_35_end_0 = const()[name = string("cache_35_end_0"), val = tensor<int32, [4]>([9, 1, 1024, 8])];
+            tensor<bool, [4]> cache_35_end_mask_0 = const()[name = string("cache_35_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_35_squeeze_mask_0 = const()[name = string("cache_35_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_35_cast_fp16 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_35_cast_fp16")];
+            tensor<int32, [1]> input_443_axes_0 = const()[name = string("input_443_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170467968)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170470080)))];
+            tensor<fp16, [1, 7, 1024]> input_443_cast_fp16 = layer_norm(axes = input_443_axes_0, beta = encoder_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_feed_forward1_weight_to_fp16, x = input_441_cast_fp16)[name = string("input_443_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170472192))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173617984))))[name = string("encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_8_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173618176)))];
+            tensor<fp16, [1, 7, 4096]> linear_73_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear1_weight_to_fp16_palettized, x = input_443_cast_fp16)[name = string("linear_73_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_447_cast_fp16 = silu(x = linear_73_cast_fp16)[name = string("input_447_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173626432))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176772224))))[name = string("encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176772416)))];
+            tensor<fp16, [1, 7, 1024]> linear_74_cast_fp16 = linear(bias = encoder_layers_8_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward1_linear2_weight_to_fp16_palettized, x = input_447_cast_fp16)[name = string("linear_74_cast_fp16")];
+            fp16 var_2159_to_fp16 = const()[name = string("op_2159_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_2160_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_2159_to_fp16)[name = string("op_2160_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_453_cast_fp16 = add(x = input_441_cast_fp16, y = var_2160_cast_fp16)[name = string("input_453_cast_fp16")];
+            tensor<int32, [1]> key_17_axes_0 = const()[name = string("key_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176774528)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176776640)))];
+            tensor<fp16, [1, 7, 1024]> key_17_cast_fp16 = layer_norm(axes = key_17_axes_0, beta = encoder_layers_8_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_self_att_weight_to_fp16, x = input_453_cast_fp16)[name = string("key_17_cast_fp16")];
+            bool input_455_interleave_0 = const()[name = string("input_455_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_455_cast_fp16 = concat(axis = var_68, interleave = input_455_interleave_0, values = (cache_33_cast_fp16, key_17_cast_fp16))[name = string("input_455_cast_fp16")];
+            tensor<int32, [3]> var_2182_begin_0 = const()[name = string("op_2182_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_2182_end_0 = const()[name = string("op_2182_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2182_end_mask_0 = const()[name = string("op_2182_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_2182_cast_fp16 = slice_by_index(begin = var_2182_begin_0, end = var_2182_end_0, end_mask = var_2182_end_mask_0, x = cache_33_cast_fp16)[name = string("op_2182_cast_fp16")];
+            bool var_2188_interleave_0 = const()[name = string("op_2188_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_2188_cast_fp16 = concat(axis = var_68, interleave = var_2188_interleave_0, values = (var_2182_cast_fp16, key_17_cast_fp16))[name = string("op_2188_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176778752))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177565248))))[name = string("encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177565440)))];
+            tensor<fp16, [1, 7, 1024]> linear_75_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_q_weight_to_fp16_palettized, x = key_17_cast_fp16)[name = string("linear_75_cast_fp16")];
+            tensor<int32, [4]> var_2193 = const()[name = string("op_2193"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_49_cast_fp16 = reshape(shape = var_2193, x = linear_75_cast_fp16)[name = string("q_49_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(177567552))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178354048))))[name = string("encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178354240)))];
+            tensor<fp16, [1, 49, 1024]> linear_76_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_k_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_76_cast_fp16")];
+            tensor<int32, [4]> var_2198 = const()[name = string("op_2198"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_33_cast_fp16 = reshape(shape = var_2198, x = linear_76_cast_fp16)[name = string("k_33_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178356352))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179142848))))[name = string("encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179143040)))];
+            tensor<fp16, [1, 49, 1024]> linear_77_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_v_weight_to_fp16_palettized, x = input_455_cast_fp16)[name = string("linear_77_cast_fp16")];
+            tensor<int32, [4]> var_2203 = const()[name = string("op_2203"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_17_cast_fp16 = reshape(shape = var_2203, x = linear_77_cast_fp16)[name = string("v_17_cast_fp16")];
+            tensor<int32, [4]> value_25_perm_0 = const()[name = string("value_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179145152)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2216_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_u_to_fp16)[name = string("op_2216_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179147264)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2218_cast_fp16 = add(x = q_49_cast_fp16, y = encoder_layers_8_self_attn_pos_bias_v_to_fp16)[name = string("op_2218_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_17_perm_0 = const()[name = string("q_with_bias_v_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_215_transpose_x_0 = const()[name = string("x_215_transpose_x_0"), val = bool(false)];
+            bool x_215_transpose_y_0 = const()[name = string("x_215_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_2220_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179149376))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179248768))))[name = string("op_2220_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2218_cast_fp16)[name = string("transpose_290")];
+            tensor<fp16, [1, 8, 7, 97]> x_215_cast_fp16 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = op_2220_to_fp16_quantized)[name = string("x_215_cast_fp16")];
+            tensor<int32, [8]> x_217_pad_0 = const()[name = string("x_217_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_217_mode_0 = const()[name = string("x_217_mode_0"), val = string("constant")];
+            fp16 const_183_to_fp16 = const()[name = string("const_183_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_217_cast_fp16 = pad(constant_val = const_183_to_fp16, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215_cast_fp16)[name = string("x_217_cast_fp16")];
+            tensor<int32, [4]> var_2228 = const()[name = string("op_2228"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_219_cast_fp16 = reshape(shape = var_2228, x = x_217_cast_fp16)[name = string("x_219_cast_fp16")];
+            tensor<int32, [4]> var_2232_begin_0 = const()[name = string("op_2232_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2232_end_0 = const()[name = string("op_2232_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_2232_end_mask_0 = const()[name = string("op_2232_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_2232_cast_fp16 = slice_by_index(begin = var_2232_begin_0, end = var_2232_end_0, end_mask = var_2232_end_mask_0, x = x_219_cast_fp16)[name = string("op_2232_cast_fp16")];
+            tensor<int32, [4]> var_2233 = const()[name = string("op_2233"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_33_cast_fp16 = reshape(shape = var_2233, x = var_2232_cast_fp16)[name = string("matrix_bd_33_cast_fp16")];
+            bool matrix_ac_17_transpose_x_0 = const()[name = string("matrix_ac_17_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_17_transpose_y_0 = const()[name = string("matrix_ac_17_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_113 = transpose(perm = transpose_113_perm_0, x = k_33_cast_fp16)[name = string("transpose_288")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_112 = transpose(perm = transpose_112_perm_0, x = var_2216_cast_fp16)[name = string("transpose_289")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_17_cast_fp16 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_112, y = transpose_113)[name = string("matrix_ac_17_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_35_begin_0 = const()[name = string("matrix_bd_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_35_end_0 = const()[name = string("matrix_bd_35_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_35_end_mask_0 = const()[name = string("matrix_bd_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_35_cast_fp16 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33_cast_fp16)[name = string("matrix_bd_35_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2242_cast_fp16 = add(x = matrix_ac_17_cast_fp16, y = matrix_bd_35_cast_fp16)[name = string("op_2242_cast_fp16")];
+            fp16 _inversed_scores_33_y_0_to_fp16 = const()[name = string("_inversed_scores_33_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_33_cast_fp16 = mul(x = var_2242_cast_fp16, y = _inversed_scores_33_y_0_to_fp16)[name = string("_inversed_scores_33_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_35_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_33_cast_fp16, cond = mask_11)[name = string("scores_35_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2248_cast_fp16 = softmax(axis = var_59, x = scores_35_cast_fp16)[name = string("op_2248_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_457_cast_fp16 = select(a = var_44_to_fp16, b = var_2248_cast_fp16, cond = mask_11)[name = string("input_457_cast_fp16")];
+            bool x_221_transpose_x_0 = const()[name = string("x_221_transpose_x_0"), val = bool(false)];
+            bool x_221_transpose_y_0 = const()[name = string("x_221_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_25_cast_fp16 = transpose(perm = value_25_perm_0, x = v_17_cast_fp16)[name = string("transpose_287")];
+            tensor<fp16, [1, 8, 7, 128]> x_221_cast_fp16 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457_cast_fp16, y = value_25_cast_fp16)[name = string("x_221_cast_fp16")];
+            tensor<int32, [4]> var_2252_perm_0 = const()[name = string("op_2252_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2253 = const()[name = string("op_2253"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_2252_cast_fp16 = transpose(perm = var_2252_perm_0, x = x_221_cast_fp16)[name = string("transpose_286")];
+            tensor<fp16, [1, 7, 1024]> input_459_cast_fp16 = reshape(shape = var_2253, x = var_2252_cast_fp16)[name = string("input_459_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179249088))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180035584))))[name = string("encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_8_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180035776)))];
+            tensor<fp16, [1, 7, 1024]> linear_79_cast_fp16 = linear(bias = encoder_layers_8_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_8_self_attn_linear_out_weight_to_fp16_palettized, x = input_459_cast_fp16)[name = string("linear_79_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_463_cast_fp16 = add(x = input_453_cast_fp16, y = linear_79_cast_fp16)[name = string("input_463_cast_fp16")];
+            tensor<int32, [1]> x_225_axes_0 = const()[name = string("x_225_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180037888)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180040000)))];
+            tensor<fp16, [1, 7, 1024]> x_225_cast_fp16 = layer_norm(axes = x_225_axes_0, beta = encoder_layers_8_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_conv_weight_to_fp16, x = input_463_cast_fp16)[name = string("x_225_cast_fp16")];
+            tensor<int32, [3]> input_465_perm_0 = const()[name = string("input_465_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_467_pad_type_0 = const()[name = string("input_467_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_467_strides_0 = const()[name = string("input_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_467_pad_0 = const()[name = string("input_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_467_dilations_0 = const()[name = string("input_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_467_groups_0 = const()[name = string("input_467_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180042112))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182139328))))[name = string("encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_465_cast_fp16 = transpose(perm = input_465_perm_0, x = x_225_cast_fp16)[name = string("transpose_285")];
+            tensor<fp16, [1, 2048, 7]> input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = encoder_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_465_cast_fp16)[name = string("input_467_cast_fp16")];
+            int32 x_227_split_num_splits_0 = const()[name = string("x_227_split_num_splits_0"), val = int32(2)];
+            int32 x_227_split_axis_0 = const()[name = string("x_227_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_227_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_227_split_cast_fp16_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467_cast_fp16)[name = string("x_227_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_227_split_1_sigmoid_cast_fp16 = sigmoid(x = x_227_split_cast_fp16_1)[name = string("x_227_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_227_cast_fp16 = mul(x = x_227_split_cast_fp16_0, y = x_227_split_1_sigmoid_cast_fp16)[name = string("x_227_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_469_cast_fp16 = select(a = var_44_to_fp16, b = x_227_cast_fp16, cond = var_575)[name = string("input_469_cast_fp16")];
+            bool new_x_35_interleave_0 = const()[name = string("new_x_35_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_35_cast_fp16 = concat(axis = var_59, interleave = new_x_35_interleave_0, values = (cache_35_cast_fp16, input_469_cast_fp16))[name = string("new_x_35_cast_fp16")];
+            tensor<int32, [3]> var_2292_begin_0 = const()[name = string("op_2292_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_2292_end_0 = const()[name = string("op_2292_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_2292_end_mask_0 = const()[name = string("op_2292_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2292_cast_fp16 = slice_by_index(begin = var_2292_begin_0, end = var_2292_end_0, end_mask = var_2292_end_mask_0, x = new_x_35_cast_fp16)[name = string("op_2292_cast_fp16")];
+            string x_229_pad_type_0 = const()[name = string("x_229_pad_type_0"), val = string("valid")];
+            int32 x_229_groups_0 = const()[name = string("x_229_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_229_strides_0 = const()[name = string("x_229_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_229_pad_0 = const()[name = string("x_229_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_229_dilations_0 = const()[name = string("x_229_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182143488))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182152768))))[name = string("encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_229_cast_fp16 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = encoder_layers_8_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_35_cast_fp16)[name = string("x_229_cast_fp16")];
+            tensor<int32, [3]> input_471_perm_0 = const()[name = string("input_471_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_231_axes_0 = const()[name = string("x_231_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182154880)))];
+            tensor<fp16, [1024]> encoder_layers_8_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_8_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182156992)))];
+            tensor<fp16, [1, 7, 1024]> input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_229_cast_fp16)[name = string("transpose_284")];
+            tensor<fp16, [1, 7, 1024]> x_231_cast_fp16 = layer_norm(axes = x_231_axes_0, beta = encoder_layers_8_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_conv_batch_norm_weight_to_fp16, x = input_471_cast_fp16)[name = string("x_231_cast_fp16")];
+            tensor<int32, [3]> input_473_perm_0 = const()[name = string("input_473_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_473_cast_fp16 = transpose(perm = input_473_perm_0, x = x_231_cast_fp16)[name = string("transpose_283")];
+            tensor<fp16, [1, 1024, 7]> input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = string("input_475_cast_fp16")];
+            string x_233_pad_type_0 = const()[name = string("x_233_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_233_strides_0 = const()[name = string("x_233_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_233_pad_0 = const()[name = string("x_233_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_233_dilations_0 = const()[name = string("x_233_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_233_groups_0 = const()[name = string("x_233_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182159104))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183207744))))[name = string("encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_233_cast_fp16 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = encoder_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_475_cast_fp16)[name = string("x_233_cast_fp16")];
+            tensor<int32, [3]> input_477_perm_0 = const()[name = string("input_477_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_477_cast_fp16 = transpose(perm = input_477_perm_0, x = x_233_cast_fp16)[name = string("transpose_282")];
+            tensor<fp16, [1, 7, 1024]> input_479_cast_fp16 = add(x = input_463_cast_fp16, y = input_477_cast_fp16)[name = string("input_479_cast_fp16")];
+            tensor<int32, [1]> input_481_axes_0 = const()[name = string("input_481_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183209856)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183211968)))];
+            tensor<fp16, [1, 7, 1024]> input_481_cast_fp16 = layer_norm(axes = input_481_axes_0, beta = encoder_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_feed_forward2_weight_to_fp16, x = input_479_cast_fp16)[name = string("input_481_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183214080))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186359872))))[name = string("encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_8_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186360064)))];
+            tensor<fp16, [1, 7, 4096]> linear_80_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear1_weight_to_fp16_palettized, x = input_481_cast_fp16)[name = string("linear_80_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_485_cast_fp16 = silu(x = linear_80_cast_fp16)[name = string("input_485_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186368320))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189514112))))[name = string("encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_8_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_8_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189514304)))];
+            tensor<fp16, [1, 7, 1024]> linear_81_cast_fp16 = linear(bias = encoder_layers_8_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_8_feed_forward2_linear2_weight_to_fp16_palettized, x = input_485_cast_fp16)[name = string("linear_81_cast_fp16")];
+            fp16 var_2335_to_fp16 = const()[name = string("op_2335_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_2336_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_2335_to_fp16)[name = string("op_2336_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_491_cast_fp16 = add(x = input_479_cast_fp16, y = var_2336_cast_fp16)[name = string("input_491_cast_fp16")];
+            tensor<int32, [1]> input_493_axes_0 = const()[name = string("input_493_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_8_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_8_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189516416)))];
+            tensor<fp16, [1024]> encoder_layers_8_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_8_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189518528)))];
+            tensor<fp16, [1, 7, 1024]> input_493_cast_fp16 = layer_norm(axes = input_493_axes_0, beta = encoder_layers_8_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_8_norm_out_weight_to_fp16, x = input_491_cast_fp16)[name = string("input_493_cast_fp16")];
+            tensor<int32, [4]> cache_37_begin_0 = const()[name = string("cache_37_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_37_end_0 = const()[name = string("cache_37_end_0"), val = tensor<int32, [4]>([10, 1, 42, 1024])];
+            tensor<bool, [4]> cache_37_end_mask_0 = const()[name = string("cache_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_37_squeeze_mask_0 = const()[name = string("cache_37_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_37_cast_fp16 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_37_cast_fp16")];
+            tensor<int32, [4]> cache_39_begin_0 = const()[name = string("cache_39_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_39_end_0 = const()[name = string("cache_39_end_0"), val = tensor<int32, [4]>([10, 1, 1024, 8])];
+            tensor<bool, [4]> cache_39_end_mask_0 = const()[name = string("cache_39_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_39_squeeze_mask_0 = const()[name = string("cache_39_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_39_cast_fp16 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_39_cast_fp16")];
+            tensor<int32, [1]> input_495_axes_0 = const()[name = string("input_495_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189520640)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189522752)))];
+            tensor<fp16, [1, 7, 1024]> input_495_cast_fp16 = layer_norm(axes = input_495_axes_0, beta = encoder_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_feed_forward1_weight_to_fp16, x = input_493_cast_fp16)[name = string("input_495_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189524864))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192670656))))[name = string("encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_9_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192670848)))];
+            tensor<fp16, [1, 7, 4096]> linear_82_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear1_weight_to_fp16_palettized, x = input_495_cast_fp16)[name = string("linear_82_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_499_cast_fp16 = silu(x = linear_82_cast_fp16)[name = string("input_499_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192679104))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195824896))))[name = string("encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195825088)))];
+            tensor<fp16, [1, 7, 1024]> linear_83_cast_fp16 = linear(bias = encoder_layers_9_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward1_linear2_weight_to_fp16_palettized, x = input_499_cast_fp16)[name = string("linear_83_cast_fp16")];
+            fp16 var_2372_to_fp16 = const()[name = string("op_2372_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_2373_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_2372_to_fp16)[name = string("op_2373_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_505_cast_fp16 = add(x = input_493_cast_fp16, y = var_2373_cast_fp16)[name = string("input_505_cast_fp16")];
+            tensor<int32, [1]> key_19_axes_0 = const()[name = string("key_19_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195827200)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195829312)))];
+            tensor<fp16, [1, 7, 1024]> key_19_cast_fp16 = layer_norm(axes = key_19_axes_0, beta = encoder_layers_9_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_self_att_weight_to_fp16, x = input_505_cast_fp16)[name = string("key_19_cast_fp16")];
+            bool input_507_interleave_0 = const()[name = string("input_507_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_507_cast_fp16 = concat(axis = var_68, interleave = input_507_interleave_0, values = (cache_37_cast_fp16, key_19_cast_fp16))[name = string("input_507_cast_fp16")];
+            tensor<int32, [3]> var_2395_begin_0 = const()[name = string("op_2395_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_2395_end_0 = const()[name = string("op_2395_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2395_end_mask_0 = const()[name = string("op_2395_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_2395_cast_fp16 = slice_by_index(begin = var_2395_begin_0, end = var_2395_end_0, end_mask = var_2395_end_mask_0, x = cache_37_cast_fp16)[name = string("op_2395_cast_fp16")];
+            bool var_2401_interleave_0 = const()[name = string("op_2401_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_2401_cast_fp16 = concat(axis = var_68, interleave = var_2401_interleave_0, values = (var_2395_cast_fp16, key_19_cast_fp16))[name = string("op_2401_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195831424))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196617920))))[name = string("encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196618112)))];
+            tensor<fp16, [1, 7, 1024]> linear_84_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_q_weight_to_fp16_palettized, x = key_19_cast_fp16)[name = string("linear_84_cast_fp16")];
+            tensor<int32, [4]> var_2406 = const()[name = string("op_2406"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_55_cast_fp16 = reshape(shape = var_2406, x = linear_84_cast_fp16)[name = string("q_55_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196620224))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197406720))))[name = string("encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197406912)))];
+            tensor<fp16, [1, 49, 1024]> linear_85_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_k_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_85_cast_fp16")];
+            tensor<int32, [4]> var_2411 = const()[name = string("op_2411"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_37_cast_fp16 = reshape(shape = var_2411, x = linear_85_cast_fp16)[name = string("k_37_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197409024))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198195520))))[name = string("encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198195712)))];
+            tensor<fp16, [1, 49, 1024]> linear_86_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_v_weight_to_fp16_palettized, x = input_507_cast_fp16)[name = string("linear_86_cast_fp16")];
+            tensor<int32, [4]> var_2416 = const()[name = string("op_2416"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_19_cast_fp16 = reshape(shape = var_2416, x = linear_86_cast_fp16)[name = string("v_19_cast_fp16")];
+            tensor<int32, [4]> value_27_perm_0 = const()[name = string("value_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198197824)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2429_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_u_to_fp16)[name = string("op_2429_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198199936)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2431_cast_fp16 = add(x = q_55_cast_fp16, y = encoder_layers_9_self_attn_pos_bias_v_to_fp16)[name = string("op_2431_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_19_perm_0 = const()[name = string("q_with_bias_v_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_241_transpose_x_0 = const()[name = string("x_241_transpose_x_0"), val = bool(false)];
+            bool x_241_transpose_y_0 = const()[name = string("x_241_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_2433_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198202048))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198301440))))[name = string("op_2433_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2431_cast_fp16)[name = string("transpose_281")];
+            tensor<fp16, [1, 8, 7, 97]> x_241_cast_fp16 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = op_2433_to_fp16_quantized)[name = string("x_241_cast_fp16")];
+            tensor<int32, [8]> x_243_pad_0 = const()[name = string("x_243_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_243_mode_0 = const()[name = string("x_243_mode_0"), val = string("constant")];
+            fp16 const_196_to_fp16 = const()[name = string("const_196_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_243_cast_fp16 = pad(constant_val = const_196_to_fp16, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241_cast_fp16)[name = string("x_243_cast_fp16")];
+            tensor<int32, [4]> var_2441 = const()[name = string("op_2441"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_245_cast_fp16 = reshape(shape = var_2441, x = x_243_cast_fp16)[name = string("x_245_cast_fp16")];
+            tensor<int32, [4]> var_2445_begin_0 = const()[name = string("op_2445_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2445_end_0 = const()[name = string("op_2445_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_2445_end_mask_0 = const()[name = string("op_2445_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_2445_cast_fp16 = slice_by_index(begin = var_2445_begin_0, end = var_2445_end_0, end_mask = var_2445_end_mask_0, x = x_245_cast_fp16)[name = string("op_2445_cast_fp16")];
+            tensor<int32, [4]> var_2446 = const()[name = string("op_2446"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_37_cast_fp16 = reshape(shape = var_2446, x = var_2445_cast_fp16)[name = string("matrix_bd_37_cast_fp16")];
+            bool matrix_ac_19_transpose_x_0 = const()[name = string("matrix_ac_19_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_19_transpose_y_0 = const()[name = string("matrix_ac_19_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_115 = transpose(perm = transpose_115_perm_0, x = k_37_cast_fp16)[name = string("transpose_279")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_114 = transpose(perm = transpose_114_perm_0, x = var_2429_cast_fp16)[name = string("transpose_280")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_19_cast_fp16 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_114, y = transpose_115)[name = string("matrix_ac_19_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_39_begin_0 = const()[name = string("matrix_bd_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_39_end_0 = const()[name = string("matrix_bd_39_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_39_end_mask_0 = const()[name = string("matrix_bd_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_39_cast_fp16 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37_cast_fp16)[name = string("matrix_bd_39_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2455_cast_fp16 = add(x = matrix_ac_19_cast_fp16, y = matrix_bd_39_cast_fp16)[name = string("op_2455_cast_fp16")];
+            fp16 _inversed_scores_37_y_0_to_fp16 = const()[name = string("_inversed_scores_37_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_37_cast_fp16 = mul(x = var_2455_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = string("_inversed_scores_37_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_39_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_37_cast_fp16, cond = mask_11)[name = string("scores_39_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2461_cast_fp16 = softmax(axis = var_59, x = scores_39_cast_fp16)[name = string("op_2461_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_509_cast_fp16 = select(a = var_44_to_fp16, b = var_2461_cast_fp16, cond = mask_11)[name = string("input_509_cast_fp16")];
+            bool x_247_transpose_x_0 = const()[name = string("x_247_transpose_x_0"), val = bool(false)];
+            bool x_247_transpose_y_0 = const()[name = string("x_247_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_27_cast_fp16 = transpose(perm = value_27_perm_0, x = v_19_cast_fp16)[name = string("transpose_278")];
+            tensor<fp16, [1, 8, 7, 128]> x_247_cast_fp16 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509_cast_fp16, y = value_27_cast_fp16)[name = string("x_247_cast_fp16")];
+            tensor<int32, [4]> var_2465_perm_0 = const()[name = string("op_2465_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2466 = const()[name = string("op_2466"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_2465_cast_fp16 = transpose(perm = var_2465_perm_0, x = x_247_cast_fp16)[name = string("transpose_277")];
+            tensor<fp16, [1, 7, 1024]> input_511_cast_fp16 = reshape(shape = var_2466, x = var_2465_cast_fp16)[name = string("input_511_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198301760))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199088256))))[name = string("encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_9_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199088448)))];
+            tensor<fp16, [1, 7, 1024]> linear_88_cast_fp16 = linear(bias = encoder_layers_9_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_9_self_attn_linear_out_weight_to_fp16_palettized, x = input_511_cast_fp16)[name = string("linear_88_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_515_cast_fp16 = add(x = input_505_cast_fp16, y = linear_88_cast_fp16)[name = string("input_515_cast_fp16")];
+            tensor<int32, [1]> x_251_axes_0 = const()[name = string("x_251_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199090560)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199092672)))];
+            tensor<fp16, [1, 7, 1024]> x_251_cast_fp16 = layer_norm(axes = x_251_axes_0, beta = encoder_layers_9_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_conv_weight_to_fp16, x = input_515_cast_fp16)[name = string("x_251_cast_fp16")];
+            tensor<int32, [3]> input_517_perm_0 = const()[name = string("input_517_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_519_pad_type_0 = const()[name = string("input_519_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_519_strides_0 = const()[name = string("input_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_519_pad_0 = const()[name = string("input_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_519_dilations_0 = const()[name = string("input_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_519_groups_0 = const()[name = string("input_519_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199094784))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201192000))))[name = string("encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_517_cast_fp16 = transpose(perm = input_517_perm_0, x = x_251_cast_fp16)[name = string("transpose_276")];
+            tensor<fp16, [1, 2048, 7]> input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = encoder_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_517_cast_fp16)[name = string("input_519_cast_fp16")];
+            int32 x_253_split_num_splits_0 = const()[name = string("x_253_split_num_splits_0"), val = int32(2)];
+            int32 x_253_split_axis_0 = const()[name = string("x_253_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_253_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_253_split_cast_fp16_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519_cast_fp16)[name = string("x_253_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_253_split_1_sigmoid_cast_fp16 = sigmoid(x = x_253_split_cast_fp16_1)[name = string("x_253_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_253_cast_fp16 = mul(x = x_253_split_cast_fp16_0, y = x_253_split_1_sigmoid_cast_fp16)[name = string("x_253_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_521_cast_fp16 = select(a = var_44_to_fp16, b = x_253_cast_fp16, cond = var_575)[name = string("input_521_cast_fp16")];
+            bool new_x_39_interleave_0 = const()[name = string("new_x_39_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_39_cast_fp16 = concat(axis = var_59, interleave = new_x_39_interleave_0, values = (cache_39_cast_fp16, input_521_cast_fp16))[name = string("new_x_39_cast_fp16")];
+            tensor<int32, [3]> var_2505_begin_0 = const()[name = string("op_2505_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_2505_end_0 = const()[name = string("op_2505_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_2505_end_mask_0 = const()[name = string("op_2505_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2505_cast_fp16 = slice_by_index(begin = var_2505_begin_0, end = var_2505_end_0, end_mask = var_2505_end_mask_0, x = new_x_39_cast_fp16)[name = string("op_2505_cast_fp16")];
+            string x_255_pad_type_0 = const()[name = string("x_255_pad_type_0"), val = string("valid")];
+            int32 x_255_groups_0 = const()[name = string("x_255_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_255_strides_0 = const()[name = string("x_255_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_255_pad_0 = const()[name = string("x_255_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_255_dilations_0 = const()[name = string("x_255_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201196160))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201205440))))[name = string("encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_255_cast_fp16 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = encoder_layers_9_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_39_cast_fp16)[name = string("x_255_cast_fp16")];
+            tensor<int32, [3]> input_523_perm_0 = const()[name = string("input_523_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_257_axes_0 = const()[name = string("x_257_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201207552)))];
+            tensor<fp16, [1024]> encoder_layers_9_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_9_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201209664)))];
+            tensor<fp16, [1, 7, 1024]> input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_255_cast_fp16)[name = string("transpose_275")];
+            tensor<fp16, [1, 7, 1024]> x_257_cast_fp16 = layer_norm(axes = x_257_axes_0, beta = encoder_layers_9_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_conv_batch_norm_weight_to_fp16, x = input_523_cast_fp16)[name = string("x_257_cast_fp16")];
+            tensor<int32, [3]> input_525_perm_0 = const()[name = string("input_525_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_525_cast_fp16 = transpose(perm = input_525_perm_0, x = x_257_cast_fp16)[name = string("transpose_274")];
+            tensor<fp16, [1, 1024, 7]> input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = string("input_527_cast_fp16")];
+            string x_259_pad_type_0 = const()[name = string("x_259_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_259_strides_0 = const()[name = string("x_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_259_pad_0 = const()[name = string("x_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_259_dilations_0 = const()[name = string("x_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_259_groups_0 = const()[name = string("x_259_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201211776))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202260416))))[name = string("encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_259_cast_fp16 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = encoder_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_527_cast_fp16)[name = string("x_259_cast_fp16")];
+            tensor<int32, [3]> input_529_perm_0 = const()[name = string("input_529_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_529_cast_fp16 = transpose(perm = input_529_perm_0, x = x_259_cast_fp16)[name = string("transpose_273")];
+            tensor<fp16, [1, 7, 1024]> input_531_cast_fp16 = add(x = input_515_cast_fp16, y = input_529_cast_fp16)[name = string("input_531_cast_fp16")];
+            tensor<int32, [1]> input_533_axes_0 = const()[name = string("input_533_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202262528)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202264640)))];
+            tensor<fp16, [1, 7, 1024]> input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_feed_forward2_weight_to_fp16, x = input_531_cast_fp16)[name = string("input_533_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(202266752))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205412544))))[name = string("encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_9_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205412736)))];
+            tensor<fp16, [1, 7, 4096]> linear_89_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear1_weight_to_fp16_palettized, x = input_533_cast_fp16)[name = string("linear_89_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_537_cast_fp16 = silu(x = linear_89_cast_fp16)[name = string("input_537_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205420992))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208566784))))[name = string("encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_9_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_9_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208566976)))];
+            tensor<fp16, [1, 7, 1024]> linear_90_cast_fp16 = linear(bias = encoder_layers_9_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_9_feed_forward2_linear2_weight_to_fp16_palettized, x = input_537_cast_fp16)[name = string("linear_90_cast_fp16")];
+            fp16 var_2548_to_fp16 = const()[name = string("op_2548_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_2549_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_2548_to_fp16)[name = string("op_2549_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_543_cast_fp16 = add(x = input_531_cast_fp16, y = var_2549_cast_fp16)[name = string("input_543_cast_fp16")];
+            tensor<int32, [1]> input_545_axes_0 = const()[name = string("input_545_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_9_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_9_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208569088)))];
+            tensor<fp16, [1024]> encoder_layers_9_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_9_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208571200)))];
+            tensor<fp16, [1, 7, 1024]> input_545_cast_fp16 = layer_norm(axes = input_545_axes_0, beta = encoder_layers_9_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_9_norm_out_weight_to_fp16, x = input_543_cast_fp16)[name = string("input_545_cast_fp16")];
+            tensor<int32, [4]> cache_41_begin_0 = const()[name = string("cache_41_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_41_end_0 = const()[name = string("cache_41_end_0"), val = tensor<int32, [4]>([11, 1, 42, 1024])];
+            tensor<bool, [4]> cache_41_end_mask_0 = const()[name = string("cache_41_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_41_squeeze_mask_0 = const()[name = string("cache_41_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_41_cast_fp16 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_41_cast_fp16")];
+            tensor<int32, [4]> cache_43_begin_0 = const()[name = string("cache_43_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_43_end_0 = const()[name = string("cache_43_end_0"), val = tensor<int32, [4]>([11, 1, 1024, 8])];
+            tensor<bool, [4]> cache_43_end_mask_0 = const()[name = string("cache_43_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_43_squeeze_mask_0 = const()[name = string("cache_43_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_43_cast_fp16 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_43_cast_fp16")];
+            tensor<int32, [1]> input_547_axes_0 = const()[name = string("input_547_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208573312)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208575424)))];
+            tensor<fp16, [1, 7, 1024]> input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_feed_forward1_weight_to_fp16, x = input_545_cast_fp16)[name = string("input_547_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208577536))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211723328))))[name = string("encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_10_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211723520)))];
+            tensor<fp16, [1, 7, 4096]> linear_91_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear1_weight_to_fp16_palettized, x = input_547_cast_fp16)[name = string("linear_91_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_551_cast_fp16 = silu(x = linear_91_cast_fp16)[name = string("input_551_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211731776))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214877568))))[name = string("encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214877760)))];
+            tensor<fp16, [1, 7, 1024]> linear_92_cast_fp16 = linear(bias = encoder_layers_10_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward1_linear2_weight_to_fp16_palettized, x = input_551_cast_fp16)[name = string("linear_92_cast_fp16")];
+            fp16 var_2585_to_fp16 = const()[name = string("op_2585_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_2586_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_2585_to_fp16)[name = string("op_2586_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_557_cast_fp16 = add(x = input_545_cast_fp16, y = var_2586_cast_fp16)[name = string("input_557_cast_fp16")];
+            tensor<int32, [1]> key_21_axes_0 = const()[name = string("key_21_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214879872)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214881984)))];
+            tensor<fp16, [1, 7, 1024]> key_21_cast_fp16 = layer_norm(axes = key_21_axes_0, beta = encoder_layers_10_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_self_att_weight_to_fp16, x = input_557_cast_fp16)[name = string("key_21_cast_fp16")];
+            bool input_559_interleave_0 = const()[name = string("input_559_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_559_cast_fp16 = concat(axis = var_68, interleave = input_559_interleave_0, values = (cache_41_cast_fp16, key_21_cast_fp16))[name = string("input_559_cast_fp16")];
+            tensor<int32, [3]> var_2608_begin_0 = const()[name = string("op_2608_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_2608_end_0 = const()[name = string("op_2608_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2608_end_mask_0 = const()[name = string("op_2608_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_2608_cast_fp16 = slice_by_index(begin = var_2608_begin_0, end = var_2608_end_0, end_mask = var_2608_end_mask_0, x = cache_41_cast_fp16)[name = string("op_2608_cast_fp16")];
+            bool var_2614_interleave_0 = const()[name = string("op_2614_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_2614_cast_fp16 = concat(axis = var_68, interleave = var_2614_interleave_0, values = (var_2608_cast_fp16, key_21_cast_fp16))[name = string("op_2614_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(214884096))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215670592))))[name = string("encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215670784)))];
+            tensor<fp16, [1, 7, 1024]> linear_93_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_q_weight_to_fp16_palettized, x = key_21_cast_fp16)[name = string("linear_93_cast_fp16")];
+            tensor<int32, [4]> var_2619 = const()[name = string("op_2619"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_61_cast_fp16 = reshape(shape = var_2619, x = linear_93_cast_fp16)[name = string("q_61_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(215672896))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216459392))))[name = string("encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216459584)))];
+            tensor<fp16, [1, 49, 1024]> linear_94_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_k_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_94_cast_fp16")];
+            tensor<int32, [4]> var_2624 = const()[name = string("op_2624"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_41_cast_fp16 = reshape(shape = var_2624, x = linear_94_cast_fp16)[name = string("k_41_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216461696))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217248192))))[name = string("encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217248384)))];
+            tensor<fp16, [1, 49, 1024]> linear_95_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_v_weight_to_fp16_palettized, x = input_559_cast_fp16)[name = string("linear_95_cast_fp16")];
+            tensor<int32, [4]> var_2629 = const()[name = string("op_2629"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_21_cast_fp16 = reshape(shape = var_2629, x = linear_95_cast_fp16)[name = string("v_21_cast_fp16")];
+            tensor<int32, [4]> value_29_perm_0 = const()[name = string("value_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217250496)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2642_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_u_to_fp16)[name = string("op_2642_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217252608)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2644_cast_fp16 = add(x = q_61_cast_fp16, y = encoder_layers_10_self_attn_pos_bias_v_to_fp16)[name = string("op_2644_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_21_perm_0 = const()[name = string("q_with_bias_v_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_267_transpose_x_0 = const()[name = string("x_267_transpose_x_0"), val = bool(false)];
+            bool x_267_transpose_y_0 = const()[name = string("x_267_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_2646_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217254720))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217354112))))[name = string("op_2646_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2644_cast_fp16)[name = string("transpose_272")];
+            tensor<fp16, [1, 8, 7, 97]> x_267_cast_fp16 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = op_2646_to_fp16_quantized)[name = string("x_267_cast_fp16")];
+            tensor<int32, [8]> x_269_pad_0 = const()[name = string("x_269_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_269_mode_0 = const()[name = string("x_269_mode_0"), val = string("constant")];
+            fp16 const_209_to_fp16 = const()[name = string("const_209_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_269_cast_fp16 = pad(constant_val = const_209_to_fp16, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267_cast_fp16)[name = string("x_269_cast_fp16")];
+            tensor<int32, [4]> var_2654 = const()[name = string("op_2654"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_271_cast_fp16 = reshape(shape = var_2654, x = x_269_cast_fp16)[name = string("x_271_cast_fp16")];
+            tensor<int32, [4]> var_2658_begin_0 = const()[name = string("op_2658_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2658_end_0 = const()[name = string("op_2658_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_2658_end_mask_0 = const()[name = string("op_2658_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_2658_cast_fp16 = slice_by_index(begin = var_2658_begin_0, end = var_2658_end_0, end_mask = var_2658_end_mask_0, x = x_271_cast_fp16)[name = string("op_2658_cast_fp16")];
+            tensor<int32, [4]> var_2659 = const()[name = string("op_2659"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_41_cast_fp16 = reshape(shape = var_2659, x = var_2658_cast_fp16)[name = string("matrix_bd_41_cast_fp16")];
+            bool matrix_ac_21_transpose_x_0 = const()[name = string("matrix_ac_21_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_21_transpose_y_0 = const()[name = string("matrix_ac_21_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_117 = transpose(perm = transpose_117_perm_0, x = k_41_cast_fp16)[name = string("transpose_270")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_116 = transpose(perm = transpose_116_perm_0, x = var_2642_cast_fp16)[name = string("transpose_271")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_21_cast_fp16 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_116, y = transpose_117)[name = string("matrix_ac_21_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_43_begin_0 = const()[name = string("matrix_bd_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_43_end_0 = const()[name = string("matrix_bd_43_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_43_end_mask_0 = const()[name = string("matrix_bd_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_43_cast_fp16 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41_cast_fp16)[name = string("matrix_bd_43_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2668_cast_fp16 = add(x = matrix_ac_21_cast_fp16, y = matrix_bd_43_cast_fp16)[name = string("op_2668_cast_fp16")];
+            fp16 _inversed_scores_41_y_0_to_fp16 = const()[name = string("_inversed_scores_41_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_41_cast_fp16 = mul(x = var_2668_cast_fp16, y = _inversed_scores_41_y_0_to_fp16)[name = string("_inversed_scores_41_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_43_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_41_cast_fp16, cond = mask_11)[name = string("scores_43_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2674_cast_fp16 = softmax(axis = var_59, x = scores_43_cast_fp16)[name = string("op_2674_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_561_cast_fp16 = select(a = var_44_to_fp16, b = var_2674_cast_fp16, cond = mask_11)[name = string("input_561_cast_fp16")];
+            bool x_273_transpose_x_0 = const()[name = string("x_273_transpose_x_0"), val = bool(false)];
+            bool x_273_transpose_y_0 = const()[name = string("x_273_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_29_cast_fp16 = transpose(perm = value_29_perm_0, x = v_21_cast_fp16)[name = string("transpose_269")];
+            tensor<fp16, [1, 8, 7, 128]> x_273_cast_fp16 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561_cast_fp16, y = value_29_cast_fp16)[name = string("x_273_cast_fp16")];
+            tensor<int32, [4]> var_2678_perm_0 = const()[name = string("op_2678_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2679 = const()[name = string("op_2679"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_2678_cast_fp16 = transpose(perm = var_2678_perm_0, x = x_273_cast_fp16)[name = string("transpose_268")];
+            tensor<fp16, [1, 7, 1024]> input_563_cast_fp16 = reshape(shape = var_2679, x = var_2678_cast_fp16)[name = string("input_563_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(217354432))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218140928))))[name = string("encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_10_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218141120)))];
+            tensor<fp16, [1, 7, 1024]> linear_97_cast_fp16 = linear(bias = encoder_layers_10_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_10_self_attn_linear_out_weight_to_fp16_palettized, x = input_563_cast_fp16)[name = string("linear_97_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_567_cast_fp16 = add(x = input_557_cast_fp16, y = linear_97_cast_fp16)[name = string("input_567_cast_fp16")];
+            tensor<int32, [1]> x_277_axes_0 = const()[name = string("x_277_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218143232)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218145344)))];
+            tensor<fp16, [1, 7, 1024]> x_277_cast_fp16 = layer_norm(axes = x_277_axes_0, beta = encoder_layers_10_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_conv_weight_to_fp16, x = input_567_cast_fp16)[name = string("x_277_cast_fp16")];
+            tensor<int32, [3]> input_569_perm_0 = const()[name = string("input_569_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_571_pad_type_0 = const()[name = string("input_571_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_571_strides_0 = const()[name = string("input_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_571_pad_0 = const()[name = string("input_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_571_dilations_0 = const()[name = string("input_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_571_groups_0 = const()[name = string("input_571_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218147456))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220244672))))[name = string("encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_569_cast_fp16 = transpose(perm = input_569_perm_0, x = x_277_cast_fp16)[name = string("transpose_267")];
+            tensor<fp16, [1, 2048, 7]> input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = encoder_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_569_cast_fp16)[name = string("input_571_cast_fp16")];
+            int32 x_279_split_num_splits_0 = const()[name = string("x_279_split_num_splits_0"), val = int32(2)];
+            int32 x_279_split_axis_0 = const()[name = string("x_279_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_279_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_279_split_cast_fp16_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571_cast_fp16)[name = string("x_279_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_279_split_1_sigmoid_cast_fp16 = sigmoid(x = x_279_split_cast_fp16_1)[name = string("x_279_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_279_cast_fp16 = mul(x = x_279_split_cast_fp16_0, y = x_279_split_1_sigmoid_cast_fp16)[name = string("x_279_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_573_cast_fp16 = select(a = var_44_to_fp16, b = x_279_cast_fp16, cond = var_575)[name = string("input_573_cast_fp16")];
+            bool new_x_43_interleave_0 = const()[name = string("new_x_43_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_43_cast_fp16 = concat(axis = var_59, interleave = new_x_43_interleave_0, values = (cache_43_cast_fp16, input_573_cast_fp16))[name = string("new_x_43_cast_fp16")];
+            tensor<int32, [3]> var_2718_begin_0 = const()[name = string("op_2718_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_2718_end_0 = const()[name = string("op_2718_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_2718_end_mask_0 = const()[name = string("op_2718_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2718_cast_fp16 = slice_by_index(begin = var_2718_begin_0, end = var_2718_end_0, end_mask = var_2718_end_mask_0, x = new_x_43_cast_fp16)[name = string("op_2718_cast_fp16")];
+            string x_281_pad_type_0 = const()[name = string("x_281_pad_type_0"), val = string("valid")];
+            int32 x_281_groups_0 = const()[name = string("x_281_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_281_strides_0 = const()[name = string("x_281_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_281_pad_0 = const()[name = string("x_281_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_281_dilations_0 = const()[name = string("x_281_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220248832))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220258112))))[name = string("encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_281_cast_fp16 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = encoder_layers_10_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_43_cast_fp16)[name = string("x_281_cast_fp16")];
+            tensor<int32, [3]> input_575_perm_0 = const()[name = string("input_575_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_283_axes_0 = const()[name = string("x_283_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220260224)))];
+            tensor<fp16, [1024]> encoder_layers_10_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_10_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220262336)))];
+            tensor<fp16, [1, 7, 1024]> input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_281_cast_fp16)[name = string("transpose_266")];
+            tensor<fp16, [1, 7, 1024]> x_283_cast_fp16 = layer_norm(axes = x_283_axes_0, beta = encoder_layers_10_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_conv_batch_norm_weight_to_fp16, x = input_575_cast_fp16)[name = string("x_283_cast_fp16")];
+            tensor<int32, [3]> input_577_perm_0 = const()[name = string("input_577_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_577_cast_fp16 = transpose(perm = input_577_perm_0, x = x_283_cast_fp16)[name = string("transpose_265")];
+            tensor<fp16, [1, 1024, 7]> input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = string("input_579_cast_fp16")];
+            string x_285_pad_type_0 = const()[name = string("x_285_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_285_strides_0 = const()[name = string("x_285_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_285_pad_0 = const()[name = string("x_285_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_285_dilations_0 = const()[name = string("x_285_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_285_groups_0 = const()[name = string("x_285_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220264448))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221313088))))[name = string("encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = encoder_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_579_cast_fp16)[name = string("x_285_cast_fp16")];
+            tensor<int32, [3]> input_581_perm_0 = const()[name = string("input_581_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_581_cast_fp16 = transpose(perm = input_581_perm_0, x = x_285_cast_fp16)[name = string("transpose_264")];
+            tensor<fp16, [1, 7, 1024]> input_583_cast_fp16 = add(x = input_567_cast_fp16, y = input_581_cast_fp16)[name = string("input_583_cast_fp16")];
+            tensor<int32, [1]> input_585_axes_0 = const()[name = string("input_585_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221315200)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221317312)))];
+            tensor<fp16, [1, 7, 1024]> input_585_cast_fp16 = layer_norm(axes = input_585_axes_0, beta = encoder_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_feed_forward2_weight_to_fp16, x = input_583_cast_fp16)[name = string("input_585_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221319424))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224465216))))[name = string("encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_10_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224465408)))];
+            tensor<fp16, [1, 7, 4096]> linear_98_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear1_weight_to_fp16_palettized, x = input_585_cast_fp16)[name = string("linear_98_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_589_cast_fp16 = silu(x = linear_98_cast_fp16)[name = string("input_589_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224473664))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227619456))))[name = string("encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_10_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_10_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227619648)))];
+            tensor<fp16, [1, 7, 1024]> linear_99_cast_fp16 = linear(bias = encoder_layers_10_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_10_feed_forward2_linear2_weight_to_fp16_palettized, x = input_589_cast_fp16)[name = string("linear_99_cast_fp16")];
+            fp16 var_2761_to_fp16 = const()[name = string("op_2761_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_2762_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_2761_to_fp16)[name = string("op_2762_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_595_cast_fp16 = add(x = input_583_cast_fp16, y = var_2762_cast_fp16)[name = string("input_595_cast_fp16")];
+            tensor<int32, [1]> input_597_axes_0 = const()[name = string("input_597_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_10_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_10_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227621760)))];
+            tensor<fp16, [1024]> encoder_layers_10_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_10_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227623872)))];
+            tensor<fp16, [1, 7, 1024]> input_597_cast_fp16 = layer_norm(axes = input_597_axes_0, beta = encoder_layers_10_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_10_norm_out_weight_to_fp16, x = input_595_cast_fp16)[name = string("input_597_cast_fp16")];
+            tensor<int32, [4]> cache_45_begin_0 = const()[name = string("cache_45_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_45_end_0 = const()[name = string("cache_45_end_0"), val = tensor<int32, [4]>([12, 1, 42, 1024])];
+            tensor<bool, [4]> cache_45_end_mask_0 = const()[name = string("cache_45_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_45_squeeze_mask_0 = const()[name = string("cache_45_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_45_cast_fp16 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_45_cast_fp16")];
+            tensor<int32, [4]> cache_47_begin_0 = const()[name = string("cache_47_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_47_end_0 = const()[name = string("cache_47_end_0"), val = tensor<int32, [4]>([12, 1, 1024, 8])];
+            tensor<bool, [4]> cache_47_end_mask_0 = const()[name = string("cache_47_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_47_squeeze_mask_0 = const()[name = string("cache_47_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_47_cast_fp16 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_47_cast_fp16")];
+            tensor<int32, [1]> input_599_axes_0 = const()[name = string("input_599_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227625984)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227628096)))];
+            tensor<fp16, [1, 7, 1024]> input_599_cast_fp16 = layer_norm(axes = input_599_axes_0, beta = encoder_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_feed_forward1_weight_to_fp16, x = input_597_cast_fp16)[name = string("input_599_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(227630208))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230776000))))[name = string("encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_11_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230776192)))];
+            tensor<fp16, [1, 7, 4096]> linear_100_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear1_weight_to_fp16_palettized, x = input_599_cast_fp16)[name = string("linear_100_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_603_cast_fp16 = silu(x = linear_100_cast_fp16)[name = string("input_603_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230784448))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233930240))))[name = string("encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233930432)))];
+            tensor<fp16, [1, 7, 1024]> linear_101_cast_fp16 = linear(bias = encoder_layers_11_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward1_linear2_weight_to_fp16_palettized, x = input_603_cast_fp16)[name = string("linear_101_cast_fp16")];
+            fp16 var_2798_to_fp16 = const()[name = string("op_2798_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_2799_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2798_to_fp16)[name = string("op_2799_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_609_cast_fp16 = add(x = input_597_cast_fp16, y = var_2799_cast_fp16)[name = string("input_609_cast_fp16")];
+            tensor<int32, [1]> key_23_axes_0 = const()[name = string("key_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233932544)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233934656)))];
+            tensor<fp16, [1, 7, 1024]> key_23_cast_fp16 = layer_norm(axes = key_23_axes_0, beta = encoder_layers_11_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_self_att_weight_to_fp16, x = input_609_cast_fp16)[name = string("key_23_cast_fp16")];
+            bool input_611_interleave_0 = const()[name = string("input_611_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_611_cast_fp16 = concat(axis = var_68, interleave = input_611_interleave_0, values = (cache_45_cast_fp16, key_23_cast_fp16))[name = string("input_611_cast_fp16")];
+            tensor<int32, [3]> var_2821_begin_0 = const()[name = string("op_2821_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_2821_end_0 = const()[name = string("op_2821_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_2821_end_mask_0 = const()[name = string("op_2821_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_2821_cast_fp16 = slice_by_index(begin = var_2821_begin_0, end = var_2821_end_0, end_mask = var_2821_end_mask_0, x = cache_45_cast_fp16)[name = string("op_2821_cast_fp16")];
+            bool var_2827_interleave_0 = const()[name = string("op_2827_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_2827_cast_fp16 = concat(axis = var_68, interleave = var_2827_interleave_0, values = (var_2821_cast_fp16, key_23_cast_fp16))[name = string("op_2827_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233936768))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234723264))))[name = string("encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234723456)))];
+            tensor<fp16, [1, 7, 1024]> linear_102_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_q_weight_to_fp16_palettized, x = key_23_cast_fp16)[name = string("linear_102_cast_fp16")];
+            tensor<int32, [4]> var_2832 = const()[name = string("op_2832"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_67_cast_fp16 = reshape(shape = var_2832, x = linear_102_cast_fp16)[name = string("q_67_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234725568))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235512064))))[name = string("encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235512256)))];
+            tensor<fp16, [1, 49, 1024]> linear_103_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_k_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_103_cast_fp16")];
+            tensor<int32, [4]> var_2837 = const()[name = string("op_2837"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_45_cast_fp16 = reshape(shape = var_2837, x = linear_103_cast_fp16)[name = string("k_45_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235514368))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236300864))))[name = string("encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236301056)))];
+            tensor<fp16, [1, 49, 1024]> linear_104_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_v_weight_to_fp16_palettized, x = input_611_cast_fp16)[name = string("linear_104_cast_fp16")];
+            tensor<int32, [4]> var_2842 = const()[name = string("op_2842"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_23_cast_fp16 = reshape(shape = var_2842, x = linear_104_cast_fp16)[name = string("v_23_cast_fp16")];
+            tensor<int32, [4]> value_31_perm_0 = const()[name = string("value_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236303168)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2855_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_u_to_fp16)[name = string("op_2855_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236305280)))];
+            tensor<fp16, [1, 7, 8, 128]> var_2857_cast_fp16 = add(x = q_67_cast_fp16, y = encoder_layers_11_self_attn_pos_bias_v_to_fp16)[name = string("op_2857_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_23_perm_0 = const()[name = string("q_with_bias_v_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_293_transpose_x_0 = const()[name = string("x_293_transpose_x_0"), val = bool(false)];
+            bool x_293_transpose_y_0 = const()[name = string("x_293_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_2859_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236307392))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236406784))))[name = string("op_2859_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2857_cast_fp16)[name = string("transpose_263")];
+            tensor<fp16, [1, 8, 7, 97]> x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = op_2859_to_fp16_quantized)[name = string("x_293_cast_fp16")];
+            tensor<int32, [8]> x_295_pad_0 = const()[name = string("x_295_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_295_mode_0 = const()[name = string("x_295_mode_0"), val = string("constant")];
+            fp16 const_222_to_fp16 = const()[name = string("const_222_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_295_cast_fp16 = pad(constant_val = const_222_to_fp16, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293_cast_fp16)[name = string("x_295_cast_fp16")];
+            tensor<int32, [4]> var_2867 = const()[name = string("op_2867"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_297_cast_fp16 = reshape(shape = var_2867, x = x_295_cast_fp16)[name = string("x_297_cast_fp16")];
+            tensor<int32, [4]> var_2871_begin_0 = const()[name = string("op_2871_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2871_end_0 = const()[name = string("op_2871_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_2871_end_mask_0 = const()[name = string("op_2871_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_2871_cast_fp16 = slice_by_index(begin = var_2871_begin_0, end = var_2871_end_0, end_mask = var_2871_end_mask_0, x = x_297_cast_fp16)[name = string("op_2871_cast_fp16")];
+            tensor<int32, [4]> var_2872 = const()[name = string("op_2872"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_45_cast_fp16 = reshape(shape = var_2872, x = var_2871_cast_fp16)[name = string("matrix_bd_45_cast_fp16")];
+            bool matrix_ac_23_transpose_x_0 = const()[name = string("matrix_ac_23_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_23_transpose_y_0 = const()[name = string("matrix_ac_23_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_119 = transpose(perm = transpose_119_perm_0, x = k_45_cast_fp16)[name = string("transpose_261")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_118 = transpose(perm = transpose_118_perm_0, x = var_2855_cast_fp16)[name = string("transpose_262")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_23_cast_fp16 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_118, y = transpose_119)[name = string("matrix_ac_23_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_47_begin_0 = const()[name = string("matrix_bd_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_47_end_0 = const()[name = string("matrix_bd_47_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_47_end_mask_0 = const()[name = string("matrix_bd_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_47_cast_fp16 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45_cast_fp16)[name = string("matrix_bd_47_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2881_cast_fp16 = add(x = matrix_ac_23_cast_fp16, y = matrix_bd_47_cast_fp16)[name = string("op_2881_cast_fp16")];
+            fp16 _inversed_scores_45_y_0_to_fp16 = const()[name = string("_inversed_scores_45_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_45_cast_fp16 = mul(x = var_2881_cast_fp16, y = _inversed_scores_45_y_0_to_fp16)[name = string("_inversed_scores_45_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_47_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_45_cast_fp16, cond = mask_11)[name = string("scores_47_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_2887_cast_fp16 = softmax(axis = var_59, x = scores_47_cast_fp16)[name = string("op_2887_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_613_cast_fp16 = select(a = var_44_to_fp16, b = var_2887_cast_fp16, cond = mask_11)[name = string("input_613_cast_fp16")];
+            bool x_299_transpose_x_0 = const()[name = string("x_299_transpose_x_0"), val = bool(false)];
+            bool x_299_transpose_y_0 = const()[name = string("x_299_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_31_cast_fp16 = transpose(perm = value_31_perm_0, x = v_23_cast_fp16)[name = string("transpose_260")];
+            tensor<fp16, [1, 8, 7, 128]> x_299_cast_fp16 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613_cast_fp16, y = value_31_cast_fp16)[name = string("x_299_cast_fp16")];
+            tensor<int32, [4]> var_2891_perm_0 = const()[name = string("op_2891_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2892 = const()[name = string("op_2892"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_2891_cast_fp16 = transpose(perm = var_2891_perm_0, x = x_299_cast_fp16)[name = string("transpose_259")];
+            tensor<fp16, [1, 7, 1024]> input_615_cast_fp16 = reshape(shape = var_2892, x = var_2891_cast_fp16)[name = string("input_615_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236407104))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237193600))))[name = string("encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_11_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237193792)))];
+            tensor<fp16, [1, 7, 1024]> linear_106_cast_fp16 = linear(bias = encoder_layers_11_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_11_self_attn_linear_out_weight_to_fp16_palettized, x = input_615_cast_fp16)[name = string("linear_106_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_619_cast_fp16 = add(x = input_609_cast_fp16, y = linear_106_cast_fp16)[name = string("input_619_cast_fp16")];
+            tensor<int32, [1]> x_303_axes_0 = const()[name = string("x_303_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237195904)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237198016)))];
+            tensor<fp16, [1, 7, 1024]> x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = encoder_layers_11_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_conv_weight_to_fp16, x = input_619_cast_fp16)[name = string("x_303_cast_fp16")];
+            tensor<int32, [3]> input_621_perm_0 = const()[name = string("input_621_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_623_pad_type_0 = const()[name = string("input_623_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_623_strides_0 = const()[name = string("input_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_623_pad_0 = const()[name = string("input_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_623_dilations_0 = const()[name = string("input_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_623_groups_0 = const()[name = string("input_623_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237200128))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239297344))))[name = string("encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_621_cast_fp16 = transpose(perm = input_621_perm_0, x = x_303_cast_fp16)[name = string("transpose_258")];
+            tensor<fp16, [1, 2048, 7]> input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = encoder_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_621_cast_fp16)[name = string("input_623_cast_fp16")];
+            int32 x_305_split_num_splits_0 = const()[name = string("x_305_split_num_splits_0"), val = int32(2)];
+            int32 x_305_split_axis_0 = const()[name = string("x_305_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_305_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_305_split_cast_fp16_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623_cast_fp16)[name = string("x_305_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_305_split_1_sigmoid_cast_fp16 = sigmoid(x = x_305_split_cast_fp16_1)[name = string("x_305_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_305_cast_fp16 = mul(x = x_305_split_cast_fp16_0, y = x_305_split_1_sigmoid_cast_fp16)[name = string("x_305_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_625_cast_fp16 = select(a = var_44_to_fp16, b = x_305_cast_fp16, cond = var_575)[name = string("input_625_cast_fp16")];
+            bool new_x_47_interleave_0 = const()[name = string("new_x_47_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_47_cast_fp16 = concat(axis = var_59, interleave = new_x_47_interleave_0, values = (cache_47_cast_fp16, input_625_cast_fp16))[name = string("new_x_47_cast_fp16")];
+            tensor<int32, [3]> var_2931_begin_0 = const()[name = string("op_2931_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_2931_end_0 = const()[name = string("op_2931_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_2931_end_mask_0 = const()[name = string("op_2931_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_2931_cast_fp16 = slice_by_index(begin = var_2931_begin_0, end = var_2931_end_0, end_mask = var_2931_end_mask_0, x = new_x_47_cast_fp16)[name = string("op_2931_cast_fp16")];
+            string x_307_pad_type_0 = const()[name = string("x_307_pad_type_0"), val = string("valid")];
+            int32 x_307_groups_0 = const()[name = string("x_307_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_307_strides_0 = const()[name = string("x_307_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_307_pad_0 = const()[name = string("x_307_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_307_dilations_0 = const()[name = string("x_307_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239301504))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239310784))))[name = string("encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = encoder_layers_11_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_47_cast_fp16)[name = string("x_307_cast_fp16")];
+            tensor<int32, [3]> input_627_perm_0 = const()[name = string("input_627_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_309_axes_0 = const()[name = string("x_309_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239312896)))];
+            tensor<fp16, [1024]> encoder_layers_11_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_11_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239315008)))];
+            tensor<fp16, [1, 7, 1024]> input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_307_cast_fp16)[name = string("transpose_257")];
+            tensor<fp16, [1, 7, 1024]> x_309_cast_fp16 = layer_norm(axes = x_309_axes_0, beta = encoder_layers_11_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_conv_batch_norm_weight_to_fp16, x = input_627_cast_fp16)[name = string("x_309_cast_fp16")];
+            tensor<int32, [3]> input_629_perm_0 = const()[name = string("input_629_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_629_cast_fp16 = transpose(perm = input_629_perm_0, x = x_309_cast_fp16)[name = string("transpose_256")];
+            tensor<fp16, [1, 1024, 7]> input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = string("input_631_cast_fp16")];
+            string x_311_pad_type_0 = const()[name = string("x_311_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_311_strides_0 = const()[name = string("x_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_311_pad_0 = const()[name = string("x_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_311_dilations_0 = const()[name = string("x_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_311_groups_0 = const()[name = string("x_311_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(239317120))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240365760))))[name = string("encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_311_cast_fp16 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = encoder_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_631_cast_fp16)[name = string("x_311_cast_fp16")];
+            tensor<int32, [3]> input_633_perm_0 = const()[name = string("input_633_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_633_cast_fp16 = transpose(perm = input_633_perm_0, x = x_311_cast_fp16)[name = string("transpose_255")];
+            tensor<fp16, [1, 7, 1024]> input_635_cast_fp16 = add(x = input_619_cast_fp16, y = input_633_cast_fp16)[name = string("input_635_cast_fp16")];
+            tensor<int32, [1]> input_637_axes_0 = const()[name = string("input_637_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240367872)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240369984)))];
+            tensor<fp16, [1, 7, 1024]> input_637_cast_fp16 = layer_norm(axes = input_637_axes_0, beta = encoder_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_feed_forward2_weight_to_fp16, x = input_635_cast_fp16)[name = string("input_637_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(240372096))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243517888))))[name = string("encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_11_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243518080)))];
+            tensor<fp16, [1, 7, 4096]> linear_107_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear1_weight_to_fp16_palettized, x = input_637_cast_fp16)[name = string("linear_107_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_641_cast_fp16 = silu(x = linear_107_cast_fp16)[name = string("input_641_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243526336))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246672128))))[name = string("encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_11_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_11_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246672320)))];
+            tensor<fp16, [1, 7, 1024]> linear_108_cast_fp16 = linear(bias = encoder_layers_11_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_11_feed_forward2_linear2_weight_to_fp16_palettized, x = input_641_cast_fp16)[name = string("linear_108_cast_fp16")];
+            fp16 var_2974_to_fp16 = const()[name = string("op_2974_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_2975_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2974_to_fp16)[name = string("op_2975_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_647_cast_fp16 = add(x = input_635_cast_fp16, y = var_2975_cast_fp16)[name = string("input_647_cast_fp16")];
+            tensor<int32, [1]> input_649_axes_0 = const()[name = string("input_649_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_11_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_11_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246674432)))];
+            tensor<fp16, [1024]> encoder_layers_11_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_11_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246676544)))];
+            tensor<fp16, [1, 7, 1024]> input_649_cast_fp16 = layer_norm(axes = input_649_axes_0, beta = encoder_layers_11_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_11_norm_out_weight_to_fp16, x = input_647_cast_fp16)[name = string("input_649_cast_fp16")];
+            tensor<int32, [4]> cache_49_begin_0 = const()[name = string("cache_49_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_49_end_0 = const()[name = string("cache_49_end_0"), val = tensor<int32, [4]>([13, 1, 42, 1024])];
+            tensor<bool, [4]> cache_49_end_mask_0 = const()[name = string("cache_49_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_49_squeeze_mask_0 = const()[name = string("cache_49_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_49_cast_fp16 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_49_cast_fp16")];
+            tensor<int32, [4]> cache_51_begin_0 = const()[name = string("cache_51_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_51_end_0 = const()[name = string("cache_51_end_0"), val = tensor<int32, [4]>([13, 1, 1024, 8])];
+            tensor<bool, [4]> cache_51_end_mask_0 = const()[name = string("cache_51_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_51_squeeze_mask_0 = const()[name = string("cache_51_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_51_cast_fp16 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_51_cast_fp16")];
+            tensor<int32, [1]> input_651_axes_0 = const()[name = string("input_651_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246678656)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246680768)))];
+            tensor<fp16, [1, 7, 1024]> input_651_cast_fp16 = layer_norm(axes = input_651_axes_0, beta = encoder_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_feed_forward1_weight_to_fp16, x = input_649_cast_fp16)[name = string("input_651_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246682880))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249828672))))[name = string("encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_12_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249828864)))];
+            tensor<fp16, [1, 7, 4096]> linear_109_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear1_weight_to_fp16_palettized, x = input_651_cast_fp16)[name = string("linear_109_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_655_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("input_655_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249837120))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252982912))))[name = string("encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252983104)))];
+            tensor<fp16, [1, 7, 1024]> linear_110_cast_fp16 = linear(bias = encoder_layers_12_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward1_linear2_weight_to_fp16_palettized, x = input_655_cast_fp16)[name = string("linear_110_cast_fp16")];
+            fp16 var_3011_to_fp16 = const()[name = string("op_3011_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_3012_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_3011_to_fp16)[name = string("op_3012_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_661_cast_fp16 = add(x = input_649_cast_fp16, y = var_3012_cast_fp16)[name = string("input_661_cast_fp16")];
+            tensor<int32, [1]> key_25_axes_0 = const()[name = string("key_25_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252985216)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252987328)))];
+            tensor<fp16, [1, 7, 1024]> key_25_cast_fp16 = layer_norm(axes = key_25_axes_0, beta = encoder_layers_12_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_self_att_weight_to_fp16, x = input_661_cast_fp16)[name = string("key_25_cast_fp16")];
+            bool input_663_interleave_0 = const()[name = string("input_663_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_663_cast_fp16 = concat(axis = var_68, interleave = input_663_interleave_0, values = (cache_49_cast_fp16, key_25_cast_fp16))[name = string("input_663_cast_fp16")];
+            tensor<int32, [3]> var_3034_begin_0 = const()[name = string("op_3034_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_3034_end_0 = const()[name = string("op_3034_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3034_end_mask_0 = const()[name = string("op_3034_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_3034_cast_fp16 = slice_by_index(begin = var_3034_begin_0, end = var_3034_end_0, end_mask = var_3034_end_mask_0, x = cache_49_cast_fp16)[name = string("op_3034_cast_fp16")];
+            bool var_3040_interleave_0 = const()[name = string("op_3040_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3040_cast_fp16 = concat(axis = var_68, interleave = var_3040_interleave_0, values = (var_3034_cast_fp16, key_25_cast_fp16))[name = string("op_3040_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(252989440))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253775936))))[name = string("encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253776128)))];
+            tensor<fp16, [1, 7, 1024]> linear_111_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_q_weight_to_fp16_palettized, x = key_25_cast_fp16)[name = string("linear_111_cast_fp16")];
+            tensor<int32, [4]> var_3045 = const()[name = string("op_3045"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_73_cast_fp16 = reshape(shape = var_3045, x = linear_111_cast_fp16)[name = string("q_73_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(253778240))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254564736))))[name = string("encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254564928)))];
+            tensor<fp16, [1, 49, 1024]> linear_112_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_k_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_112_cast_fp16")];
+            tensor<int32, [4]> var_3050 = const()[name = string("op_3050"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_49_cast_fp16 = reshape(shape = var_3050, x = linear_112_cast_fp16)[name = string("k_49_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254567040))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255353536))))[name = string("encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255353728)))];
+            tensor<fp16, [1, 49, 1024]> linear_113_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_v_weight_to_fp16_palettized, x = input_663_cast_fp16)[name = string("linear_113_cast_fp16")];
+            tensor<int32, [4]> var_3055 = const()[name = string("op_3055"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_25_cast_fp16 = reshape(shape = var_3055, x = linear_113_cast_fp16)[name = string("v_25_cast_fp16")];
+            tensor<int32, [4]> value_33_perm_0 = const()[name = string("value_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255355840)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3068_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_u_to_fp16)[name = string("op_3068_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255357952)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3070_cast_fp16 = add(x = q_73_cast_fp16, y = encoder_layers_12_self_attn_pos_bias_v_to_fp16)[name = string("op_3070_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_25_perm_0 = const()[name = string("q_with_bias_v_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_319_transpose_x_0 = const()[name = string("x_319_transpose_x_0"), val = bool(false)];
+            bool x_319_transpose_y_0 = const()[name = string("x_319_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_3072_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255360064))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255459456))))[name = string("op_3072_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_3070_cast_fp16)[name = string("transpose_254")];
+            tensor<fp16, [1, 8, 7, 97]> x_319_cast_fp16 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = op_3072_to_fp16_quantized)[name = string("x_319_cast_fp16")];
+            tensor<int32, [8]> x_321_pad_0 = const()[name = string("x_321_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_321_mode_0 = const()[name = string("x_321_mode_0"), val = string("constant")];
+            fp16 const_235_to_fp16 = const()[name = string("const_235_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_321_cast_fp16 = pad(constant_val = const_235_to_fp16, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319_cast_fp16)[name = string("x_321_cast_fp16")];
+            tensor<int32, [4]> var_3080 = const()[name = string("op_3080"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_323_cast_fp16 = reshape(shape = var_3080, x = x_321_cast_fp16)[name = string("x_323_cast_fp16")];
+            tensor<int32, [4]> var_3084_begin_0 = const()[name = string("op_3084_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3084_end_0 = const()[name = string("op_3084_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_3084_end_mask_0 = const()[name = string("op_3084_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_3084_cast_fp16 = slice_by_index(begin = var_3084_begin_0, end = var_3084_end_0, end_mask = var_3084_end_mask_0, x = x_323_cast_fp16)[name = string("op_3084_cast_fp16")];
+            tensor<int32, [4]> var_3085 = const()[name = string("op_3085"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_49_cast_fp16 = reshape(shape = var_3085, x = var_3084_cast_fp16)[name = string("matrix_bd_49_cast_fp16")];
+            bool matrix_ac_25_transpose_x_0 = const()[name = string("matrix_ac_25_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_25_transpose_y_0 = const()[name = string("matrix_ac_25_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_120_perm_0 = const()[name = string("transpose_120_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_121_perm_0 = const()[name = string("transpose_121_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_121 = transpose(perm = transpose_121_perm_0, x = k_49_cast_fp16)[name = string("transpose_252")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_120 = transpose(perm = transpose_120_perm_0, x = var_3068_cast_fp16)[name = string("transpose_253")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_25_cast_fp16 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_120, y = transpose_121)[name = string("matrix_ac_25_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_51_begin_0 = const()[name = string("matrix_bd_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_51_end_0 = const()[name = string("matrix_bd_51_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_51_end_mask_0 = const()[name = string("matrix_bd_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_51_cast_fp16 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49_cast_fp16)[name = string("matrix_bd_51_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3094_cast_fp16 = add(x = matrix_ac_25_cast_fp16, y = matrix_bd_51_cast_fp16)[name = string("op_3094_cast_fp16")];
+            fp16 _inversed_scores_49_y_0_to_fp16 = const()[name = string("_inversed_scores_49_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_49_cast_fp16 = mul(x = var_3094_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = string("_inversed_scores_49_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_51_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_49_cast_fp16, cond = mask_11)[name = string("scores_51_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3100_cast_fp16 = softmax(axis = var_59, x = scores_51_cast_fp16)[name = string("op_3100_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_665_cast_fp16 = select(a = var_44_to_fp16, b = var_3100_cast_fp16, cond = mask_11)[name = string("input_665_cast_fp16")];
+            bool x_325_transpose_x_0 = const()[name = string("x_325_transpose_x_0"), val = bool(false)];
+            bool x_325_transpose_y_0 = const()[name = string("x_325_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_33_cast_fp16 = transpose(perm = value_33_perm_0, x = v_25_cast_fp16)[name = string("transpose_251")];
+            tensor<fp16, [1, 8, 7, 128]> x_325_cast_fp16 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665_cast_fp16, y = value_33_cast_fp16)[name = string("x_325_cast_fp16")];
+            tensor<int32, [4]> var_3104_perm_0 = const()[name = string("op_3104_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3105 = const()[name = string("op_3105"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_3104_cast_fp16 = transpose(perm = var_3104_perm_0, x = x_325_cast_fp16)[name = string("transpose_250")];
+            tensor<fp16, [1, 7, 1024]> input_667_cast_fp16 = reshape(shape = var_3105, x = var_3104_cast_fp16)[name = string("input_667_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(255459776))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256246272))))[name = string("encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_12_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256246464)))];
+            tensor<fp16, [1, 7, 1024]> linear_115_cast_fp16 = linear(bias = encoder_layers_12_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_12_self_attn_linear_out_weight_to_fp16_palettized, x = input_667_cast_fp16)[name = string("linear_115_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_671_cast_fp16 = add(x = input_661_cast_fp16, y = linear_115_cast_fp16)[name = string("input_671_cast_fp16")];
+            tensor<int32, [1]> x_329_axes_0 = const()[name = string("x_329_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256248576)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256250688)))];
+            tensor<fp16, [1, 7, 1024]> x_329_cast_fp16 = layer_norm(axes = x_329_axes_0, beta = encoder_layers_12_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_conv_weight_to_fp16, x = input_671_cast_fp16)[name = string("x_329_cast_fp16")];
+            tensor<int32, [3]> input_673_perm_0 = const()[name = string("input_673_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_675_pad_type_0 = const()[name = string("input_675_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_675_strides_0 = const()[name = string("input_675_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_675_pad_0 = const()[name = string("input_675_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_675_dilations_0 = const()[name = string("input_675_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_675_groups_0 = const()[name = string("input_675_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256252800))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258350016))))[name = string("encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_673_cast_fp16 = transpose(perm = input_673_perm_0, x = x_329_cast_fp16)[name = string("transpose_249")];
+            tensor<fp16, [1, 2048, 7]> input_675_cast_fp16 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = encoder_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_673_cast_fp16)[name = string("input_675_cast_fp16")];
+            int32 x_331_split_num_splits_0 = const()[name = string("x_331_split_num_splits_0"), val = int32(2)];
+            int32 x_331_split_axis_0 = const()[name = string("x_331_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_331_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_331_split_cast_fp16_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675_cast_fp16)[name = string("x_331_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_331_split_1_sigmoid_cast_fp16 = sigmoid(x = x_331_split_cast_fp16_1)[name = string("x_331_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_331_cast_fp16 = mul(x = x_331_split_cast_fp16_0, y = x_331_split_1_sigmoid_cast_fp16)[name = string("x_331_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_677_cast_fp16 = select(a = var_44_to_fp16, b = x_331_cast_fp16, cond = var_575)[name = string("input_677_cast_fp16")];
+            bool new_x_51_interleave_0 = const()[name = string("new_x_51_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_51_cast_fp16 = concat(axis = var_59, interleave = new_x_51_interleave_0, values = (cache_51_cast_fp16, input_677_cast_fp16))[name = string("new_x_51_cast_fp16")];
+            tensor<int32, [3]> var_3144_begin_0 = const()[name = string("op_3144_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_3144_end_0 = const()[name = string("op_3144_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_3144_end_mask_0 = const()[name = string("op_3144_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3144_cast_fp16 = slice_by_index(begin = var_3144_begin_0, end = var_3144_end_0, end_mask = var_3144_end_mask_0, x = new_x_51_cast_fp16)[name = string("op_3144_cast_fp16")];
+            string x_333_pad_type_0 = const()[name = string("x_333_pad_type_0"), val = string("valid")];
+            int32 x_333_groups_0 = const()[name = string("x_333_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_333_strides_0 = const()[name = string("x_333_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_333_pad_0 = const()[name = string("x_333_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_333_dilations_0 = const()[name = string("x_333_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258354176))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258363456))))[name = string("encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_333_cast_fp16 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = encoder_layers_12_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_51_cast_fp16)[name = string("x_333_cast_fp16")];
+            tensor<int32, [3]> input_679_perm_0 = const()[name = string("input_679_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_335_axes_0 = const()[name = string("x_335_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258365568)))];
+            tensor<fp16, [1024]> encoder_layers_12_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_12_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258367680)))];
+            tensor<fp16, [1, 7, 1024]> input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_333_cast_fp16)[name = string("transpose_248")];
+            tensor<fp16, [1, 7, 1024]> x_335_cast_fp16 = layer_norm(axes = x_335_axes_0, beta = encoder_layers_12_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_conv_batch_norm_weight_to_fp16, x = input_679_cast_fp16)[name = string("x_335_cast_fp16")];
+            tensor<int32, [3]> input_681_perm_0 = const()[name = string("input_681_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_681_cast_fp16 = transpose(perm = input_681_perm_0, x = x_335_cast_fp16)[name = string("transpose_247")];
+            tensor<fp16, [1, 1024, 7]> input_683_cast_fp16 = silu(x = input_681_cast_fp16)[name = string("input_683_cast_fp16")];
+            string x_337_pad_type_0 = const()[name = string("x_337_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_337_strides_0 = const()[name = string("x_337_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_337_pad_0 = const()[name = string("x_337_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_337_dilations_0 = const()[name = string("x_337_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_337_groups_0 = const()[name = string("x_337_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258369792))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259418432))))[name = string("encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_337_cast_fp16 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = encoder_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_683_cast_fp16)[name = string("x_337_cast_fp16")];
+            tensor<int32, [3]> input_685_perm_0 = const()[name = string("input_685_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_685_cast_fp16 = transpose(perm = input_685_perm_0, x = x_337_cast_fp16)[name = string("transpose_246")];
+            tensor<fp16, [1, 7, 1024]> input_687_cast_fp16 = add(x = input_671_cast_fp16, y = input_685_cast_fp16)[name = string("input_687_cast_fp16")];
+            tensor<int32, [1]> input_689_axes_0 = const()[name = string("input_689_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259420544)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259422656)))];
+            tensor<fp16, [1, 7, 1024]> input_689_cast_fp16 = layer_norm(axes = input_689_axes_0, beta = encoder_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_feed_forward2_weight_to_fp16, x = input_687_cast_fp16)[name = string("input_689_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259424768))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262570560))))[name = string("encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_12_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262570752)))];
+            tensor<fp16, [1, 7, 4096]> linear_116_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear1_weight_to_fp16_palettized, x = input_689_cast_fp16)[name = string("linear_116_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_693_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("input_693_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262579008))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265724800))))[name = string("encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_12_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_12_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265724992)))];
+            tensor<fp16, [1, 7, 1024]> linear_117_cast_fp16 = linear(bias = encoder_layers_12_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_12_feed_forward2_linear2_weight_to_fp16_palettized, x = input_693_cast_fp16)[name = string("linear_117_cast_fp16")];
+            fp16 var_3187_to_fp16 = const()[name = string("op_3187_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_3188_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_3187_to_fp16)[name = string("op_3188_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_699_cast_fp16 = add(x = input_687_cast_fp16, y = var_3188_cast_fp16)[name = string("input_699_cast_fp16")];
+            tensor<int32, [1]> input_701_axes_0 = const()[name = string("input_701_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_12_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_12_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265727104)))];
+            tensor<fp16, [1024]> encoder_layers_12_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_12_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265729216)))];
+            tensor<fp16, [1, 7, 1024]> input_701_cast_fp16 = layer_norm(axes = input_701_axes_0, beta = encoder_layers_12_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_12_norm_out_weight_to_fp16, x = input_699_cast_fp16)[name = string("input_701_cast_fp16")];
+            tensor<int32, [4]> cache_53_begin_0 = const()[name = string("cache_53_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_53_end_0 = const()[name = string("cache_53_end_0"), val = tensor<int32, [4]>([14, 1, 42, 1024])];
+            tensor<bool, [4]> cache_53_end_mask_0 = const()[name = string("cache_53_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_53_squeeze_mask_0 = const()[name = string("cache_53_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_53_cast_fp16 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_53_cast_fp16")];
+            tensor<int32, [4]> cache_55_begin_0 = const()[name = string("cache_55_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_55_end_0 = const()[name = string("cache_55_end_0"), val = tensor<int32, [4]>([14, 1, 1024, 8])];
+            tensor<bool, [4]> cache_55_end_mask_0 = const()[name = string("cache_55_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_55_squeeze_mask_0 = const()[name = string("cache_55_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_55_cast_fp16 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_55_cast_fp16")];
+            tensor<int32, [1]> input_703_axes_0 = const()[name = string("input_703_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265731328)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265733440)))];
+            tensor<fp16, [1, 7, 1024]> input_703_cast_fp16 = layer_norm(axes = input_703_axes_0, beta = encoder_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_feed_forward1_weight_to_fp16, x = input_701_cast_fp16)[name = string("input_703_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(265735552))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268881344))))[name = string("encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_13_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268881536)))];
+            tensor<fp16, [1, 7, 4096]> linear_118_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear1_weight_to_fp16_palettized, x = input_703_cast_fp16)[name = string("linear_118_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_707_cast_fp16 = silu(x = linear_118_cast_fp16)[name = string("input_707_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268889792))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272035584))))[name = string("encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272035776)))];
+            tensor<fp16, [1, 7, 1024]> linear_119_cast_fp16 = linear(bias = encoder_layers_13_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward1_linear2_weight_to_fp16_palettized, x = input_707_cast_fp16)[name = string("linear_119_cast_fp16")];
+            fp16 var_3224_to_fp16 = const()[name = string("op_3224_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_3225_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_3224_to_fp16)[name = string("op_3225_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_713_cast_fp16 = add(x = input_701_cast_fp16, y = var_3225_cast_fp16)[name = string("input_713_cast_fp16")];
+            tensor<int32, [1]> key_27_axes_0 = const()[name = string("key_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272037888)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272040000)))];
+            tensor<fp16, [1, 7, 1024]> key_27_cast_fp16 = layer_norm(axes = key_27_axes_0, beta = encoder_layers_13_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_self_att_weight_to_fp16, x = input_713_cast_fp16)[name = string("key_27_cast_fp16")];
+            bool input_715_interleave_0 = const()[name = string("input_715_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_715_cast_fp16 = concat(axis = var_68, interleave = input_715_interleave_0, values = (cache_53_cast_fp16, key_27_cast_fp16))[name = string("input_715_cast_fp16")];
+            tensor<int32, [3]> var_3247_begin_0 = const()[name = string("op_3247_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_3247_end_0 = const()[name = string("op_3247_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3247_end_mask_0 = const()[name = string("op_3247_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_3247_cast_fp16 = slice_by_index(begin = var_3247_begin_0, end = var_3247_end_0, end_mask = var_3247_end_mask_0, x = cache_53_cast_fp16)[name = string("op_3247_cast_fp16")];
+            bool var_3253_interleave_0 = const()[name = string("op_3253_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3253_cast_fp16 = concat(axis = var_68, interleave = var_3253_interleave_0, values = (var_3247_cast_fp16, key_27_cast_fp16))[name = string("op_3253_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272042112))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272828608))))[name = string("encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272828800)))];
+            tensor<fp16, [1, 7, 1024]> linear_120_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_q_weight_to_fp16_palettized, x = key_27_cast_fp16)[name = string("linear_120_cast_fp16")];
+            tensor<int32, [4]> var_3258 = const()[name = string("op_3258"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_79_cast_fp16 = reshape(shape = var_3258, x = linear_120_cast_fp16)[name = string("q_79_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272830912))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273617408))))[name = string("encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273617600)))];
+            tensor<fp16, [1, 49, 1024]> linear_121_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_k_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_121_cast_fp16")];
+            tensor<int32, [4]> var_3263 = const()[name = string("op_3263"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_53_cast_fp16 = reshape(shape = var_3263, x = linear_121_cast_fp16)[name = string("k_53_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273619712))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274406208))))[name = string("encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274406400)))];
+            tensor<fp16, [1, 49, 1024]> linear_122_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_v_weight_to_fp16_palettized, x = input_715_cast_fp16)[name = string("linear_122_cast_fp16")];
+            tensor<int32, [4]> var_3268 = const()[name = string("op_3268"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_27_cast_fp16 = reshape(shape = var_3268, x = linear_122_cast_fp16)[name = string("v_27_cast_fp16")];
+            tensor<int32, [4]> value_35_perm_0 = const()[name = string("value_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274408512)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3281_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_u_to_fp16)[name = string("op_3281_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274410624)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3283_cast_fp16 = add(x = q_79_cast_fp16, y = encoder_layers_13_self_attn_pos_bias_v_to_fp16)[name = string("op_3283_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_27_perm_0 = const()[name = string("q_with_bias_v_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_345_transpose_x_0 = const()[name = string("x_345_transpose_x_0"), val = bool(false)];
+            bool x_345_transpose_y_0 = const()[name = string("x_345_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_3285_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274412736))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274512128))))[name = string("op_3285_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3283_cast_fp16)[name = string("transpose_245")];
+            tensor<fp16, [1, 8, 7, 97]> x_345_cast_fp16 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = op_3285_to_fp16_quantized)[name = string("x_345_cast_fp16")];
+            tensor<int32, [8]> x_347_pad_0 = const()[name = string("x_347_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_347_mode_0 = const()[name = string("x_347_mode_0"), val = string("constant")];
+            fp16 const_248_to_fp16 = const()[name = string("const_248_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_347_cast_fp16 = pad(constant_val = const_248_to_fp16, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345_cast_fp16)[name = string("x_347_cast_fp16")];
+            tensor<int32, [4]> var_3293 = const()[name = string("op_3293"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_349_cast_fp16 = reshape(shape = var_3293, x = x_347_cast_fp16)[name = string("x_349_cast_fp16")];
+            tensor<int32, [4]> var_3297_begin_0 = const()[name = string("op_3297_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3297_end_0 = const()[name = string("op_3297_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_3297_end_mask_0 = const()[name = string("op_3297_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_3297_cast_fp16 = slice_by_index(begin = var_3297_begin_0, end = var_3297_end_0, end_mask = var_3297_end_mask_0, x = x_349_cast_fp16)[name = string("op_3297_cast_fp16")];
+            tensor<int32, [4]> var_3298 = const()[name = string("op_3298"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_53_cast_fp16 = reshape(shape = var_3298, x = var_3297_cast_fp16)[name = string("matrix_bd_53_cast_fp16")];
+            bool matrix_ac_27_transpose_x_0 = const()[name = string("matrix_ac_27_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_27_transpose_y_0 = const()[name = string("matrix_ac_27_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_122_perm_0 = const()[name = string("transpose_122_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_123_perm_0 = const()[name = string("transpose_123_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_123 = transpose(perm = transpose_123_perm_0, x = k_53_cast_fp16)[name = string("transpose_243")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_122 = transpose(perm = transpose_122_perm_0, x = var_3281_cast_fp16)[name = string("transpose_244")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_27_cast_fp16 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_122, y = transpose_123)[name = string("matrix_ac_27_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_55_begin_0 = const()[name = string("matrix_bd_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_55_end_0 = const()[name = string("matrix_bd_55_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_55_end_mask_0 = const()[name = string("matrix_bd_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_55_cast_fp16 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53_cast_fp16)[name = string("matrix_bd_55_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3307_cast_fp16 = add(x = matrix_ac_27_cast_fp16, y = matrix_bd_55_cast_fp16)[name = string("op_3307_cast_fp16")];
+            fp16 _inversed_scores_53_y_0_to_fp16 = const()[name = string("_inversed_scores_53_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_53_cast_fp16 = mul(x = var_3307_cast_fp16, y = _inversed_scores_53_y_0_to_fp16)[name = string("_inversed_scores_53_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_55_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_53_cast_fp16, cond = mask_11)[name = string("scores_55_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3313_cast_fp16 = softmax(axis = var_59, x = scores_55_cast_fp16)[name = string("op_3313_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_717_cast_fp16 = select(a = var_44_to_fp16, b = var_3313_cast_fp16, cond = mask_11)[name = string("input_717_cast_fp16")];
+            bool x_351_transpose_x_0 = const()[name = string("x_351_transpose_x_0"), val = bool(false)];
+            bool x_351_transpose_y_0 = const()[name = string("x_351_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_35_cast_fp16 = transpose(perm = value_35_perm_0, x = v_27_cast_fp16)[name = string("transpose_242")];
+            tensor<fp16, [1, 8, 7, 128]> x_351_cast_fp16 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717_cast_fp16, y = value_35_cast_fp16)[name = string("x_351_cast_fp16")];
+            tensor<int32, [4]> var_3317_perm_0 = const()[name = string("op_3317_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3318 = const()[name = string("op_3318"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_3317_cast_fp16 = transpose(perm = var_3317_perm_0, x = x_351_cast_fp16)[name = string("transpose_241")];
+            tensor<fp16, [1, 7, 1024]> input_719_cast_fp16 = reshape(shape = var_3318, x = var_3317_cast_fp16)[name = string("input_719_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274512448))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275298944))))[name = string("encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_13_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275299136)))];
+            tensor<fp16, [1, 7, 1024]> linear_124_cast_fp16 = linear(bias = encoder_layers_13_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_13_self_attn_linear_out_weight_to_fp16_palettized, x = input_719_cast_fp16)[name = string("linear_124_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_723_cast_fp16 = add(x = input_713_cast_fp16, y = linear_124_cast_fp16)[name = string("input_723_cast_fp16")];
+            tensor<int32, [1]> x_355_axes_0 = const()[name = string("x_355_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275301248)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275303360)))];
+            tensor<fp16, [1, 7, 1024]> x_355_cast_fp16 = layer_norm(axes = x_355_axes_0, beta = encoder_layers_13_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_conv_weight_to_fp16, x = input_723_cast_fp16)[name = string("x_355_cast_fp16")];
+            tensor<int32, [3]> input_725_perm_0 = const()[name = string("input_725_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_727_pad_type_0 = const()[name = string("input_727_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_727_strides_0 = const()[name = string("input_727_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_727_pad_0 = const()[name = string("input_727_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_727_dilations_0 = const()[name = string("input_727_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_727_groups_0 = const()[name = string("input_727_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275305472))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277402688))))[name = string("encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_725_cast_fp16 = transpose(perm = input_725_perm_0, x = x_355_cast_fp16)[name = string("transpose_240")];
+            tensor<fp16, [1, 2048, 7]> input_727_cast_fp16 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = encoder_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_725_cast_fp16)[name = string("input_727_cast_fp16")];
+            int32 x_357_split_num_splits_0 = const()[name = string("x_357_split_num_splits_0"), val = int32(2)];
+            int32 x_357_split_axis_0 = const()[name = string("x_357_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_357_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_357_split_cast_fp16_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727_cast_fp16)[name = string("x_357_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_357_split_1_sigmoid_cast_fp16 = sigmoid(x = x_357_split_cast_fp16_1)[name = string("x_357_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_357_cast_fp16 = mul(x = x_357_split_cast_fp16_0, y = x_357_split_1_sigmoid_cast_fp16)[name = string("x_357_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_729_cast_fp16 = select(a = var_44_to_fp16, b = x_357_cast_fp16, cond = var_575)[name = string("input_729_cast_fp16")];
+            bool new_x_55_interleave_0 = const()[name = string("new_x_55_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_55_cast_fp16 = concat(axis = var_59, interleave = new_x_55_interleave_0, values = (cache_55_cast_fp16, input_729_cast_fp16))[name = string("new_x_55_cast_fp16")];
+            tensor<int32, [3]> var_3357_begin_0 = const()[name = string("op_3357_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_3357_end_0 = const()[name = string("op_3357_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_3357_end_mask_0 = const()[name = string("op_3357_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3357_cast_fp16 = slice_by_index(begin = var_3357_begin_0, end = var_3357_end_0, end_mask = var_3357_end_mask_0, x = new_x_55_cast_fp16)[name = string("op_3357_cast_fp16")];
+            string x_359_pad_type_0 = const()[name = string("x_359_pad_type_0"), val = string("valid")];
+            int32 x_359_groups_0 = const()[name = string("x_359_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_359_strides_0 = const()[name = string("x_359_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_359_pad_0 = const()[name = string("x_359_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_359_dilations_0 = const()[name = string("x_359_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277406848))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277416128))))[name = string("encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_359_cast_fp16 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = encoder_layers_13_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_55_cast_fp16)[name = string("x_359_cast_fp16")];
+            tensor<int32, [3]> input_731_perm_0 = const()[name = string("input_731_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_361_axes_0 = const()[name = string("x_361_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277418240)))];
+            tensor<fp16, [1024]> encoder_layers_13_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_13_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277420352)))];
+            tensor<fp16, [1, 7, 1024]> input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_359_cast_fp16)[name = string("transpose_239")];
+            tensor<fp16, [1, 7, 1024]> x_361_cast_fp16 = layer_norm(axes = x_361_axes_0, beta = encoder_layers_13_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_conv_batch_norm_weight_to_fp16, x = input_731_cast_fp16)[name = string("x_361_cast_fp16")];
+            tensor<int32, [3]> input_733_perm_0 = const()[name = string("input_733_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_733_cast_fp16 = transpose(perm = input_733_perm_0, x = x_361_cast_fp16)[name = string("transpose_238")];
+            tensor<fp16, [1, 1024, 7]> input_735_cast_fp16 = silu(x = input_733_cast_fp16)[name = string("input_735_cast_fp16")];
+            string x_363_pad_type_0 = const()[name = string("x_363_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_363_strides_0 = const()[name = string("x_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_363_pad_0 = const()[name = string("x_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_363_dilations_0 = const()[name = string("x_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_363_groups_0 = const()[name = string("x_363_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277422464))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278471104))))[name = string("encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_363_cast_fp16 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = encoder_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_735_cast_fp16)[name = string("x_363_cast_fp16")];
+            tensor<int32, [3]> input_737_perm_0 = const()[name = string("input_737_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_737_cast_fp16 = transpose(perm = input_737_perm_0, x = x_363_cast_fp16)[name = string("transpose_237")];
+            tensor<fp16, [1, 7, 1024]> input_739_cast_fp16 = add(x = input_723_cast_fp16, y = input_737_cast_fp16)[name = string("input_739_cast_fp16")];
+            tensor<int32, [1]> input_741_axes_0 = const()[name = string("input_741_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278473216)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278475328)))];
+            tensor<fp16, [1, 7, 1024]> input_741_cast_fp16 = layer_norm(axes = input_741_axes_0, beta = encoder_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_feed_forward2_weight_to_fp16, x = input_739_cast_fp16)[name = string("input_741_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(278477440))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281623232))))[name = string("encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_13_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281623424)))];
+            tensor<fp16, [1, 7, 4096]> linear_125_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear1_weight_to_fp16_palettized, x = input_741_cast_fp16)[name = string("linear_125_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_745_cast_fp16 = silu(x = linear_125_cast_fp16)[name = string("input_745_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281631680))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284777472))))[name = string("encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_13_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_13_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284777664)))];
+            tensor<fp16, [1, 7, 1024]> linear_126_cast_fp16 = linear(bias = encoder_layers_13_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_13_feed_forward2_linear2_weight_to_fp16_palettized, x = input_745_cast_fp16)[name = string("linear_126_cast_fp16")];
+            fp16 var_3400_to_fp16 = const()[name = string("op_3400_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_3401_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_3400_to_fp16)[name = string("op_3401_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_751_cast_fp16 = add(x = input_739_cast_fp16, y = var_3401_cast_fp16)[name = string("input_751_cast_fp16")];
+            tensor<int32, [1]> input_753_axes_0 = const()[name = string("input_753_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_13_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_13_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284779776)))];
+            tensor<fp16, [1024]> encoder_layers_13_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_13_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284781888)))];
+            tensor<fp16, [1, 7, 1024]> input_753_cast_fp16 = layer_norm(axes = input_753_axes_0, beta = encoder_layers_13_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_13_norm_out_weight_to_fp16, x = input_751_cast_fp16)[name = string("input_753_cast_fp16")];
+            tensor<int32, [4]> cache_57_begin_0 = const()[name = string("cache_57_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_57_end_0 = const()[name = string("cache_57_end_0"), val = tensor<int32, [4]>([15, 1, 42, 1024])];
+            tensor<bool, [4]> cache_57_end_mask_0 = const()[name = string("cache_57_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_57_squeeze_mask_0 = const()[name = string("cache_57_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_57_cast_fp16 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_57_cast_fp16")];
+            tensor<int32, [4]> cache_59_begin_0 = const()[name = string("cache_59_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_59_end_0 = const()[name = string("cache_59_end_0"), val = tensor<int32, [4]>([15, 1, 1024, 8])];
+            tensor<bool, [4]> cache_59_end_mask_0 = const()[name = string("cache_59_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_59_squeeze_mask_0 = const()[name = string("cache_59_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_59_cast_fp16 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_59_cast_fp16")];
+            tensor<int32, [1]> input_755_axes_0 = const()[name = string("input_755_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284784000)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284786112)))];
+            tensor<fp16, [1, 7, 1024]> input_755_cast_fp16 = layer_norm(axes = input_755_axes_0, beta = encoder_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_feed_forward1_weight_to_fp16, x = input_753_cast_fp16)[name = string("input_755_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284788224))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287934016))))[name = string("encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_14_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287934208)))];
+            tensor<fp16, [1, 7, 4096]> linear_127_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear1_weight_to_fp16_palettized, x = input_755_cast_fp16)[name = string("linear_127_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_759_cast_fp16 = silu(x = linear_127_cast_fp16)[name = string("input_759_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287942464))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291088256))))[name = string("encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291088448)))];
+            tensor<fp16, [1, 7, 1024]> linear_128_cast_fp16 = linear(bias = encoder_layers_14_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward1_linear2_weight_to_fp16_palettized, x = input_759_cast_fp16)[name = string("linear_128_cast_fp16")];
+            fp16 var_3437_to_fp16 = const()[name = string("op_3437_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_3438_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_3437_to_fp16)[name = string("op_3438_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_765_cast_fp16 = add(x = input_753_cast_fp16, y = var_3438_cast_fp16)[name = string("input_765_cast_fp16")];
+            tensor<int32, [1]> key_29_axes_0 = const()[name = string("key_29_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291090560)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291092672)))];
+            tensor<fp16, [1, 7, 1024]> key_29_cast_fp16 = layer_norm(axes = key_29_axes_0, beta = encoder_layers_14_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_self_att_weight_to_fp16, x = input_765_cast_fp16)[name = string("key_29_cast_fp16")];
+            bool input_767_interleave_0 = const()[name = string("input_767_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_767_cast_fp16 = concat(axis = var_68, interleave = input_767_interleave_0, values = (cache_57_cast_fp16, key_29_cast_fp16))[name = string("input_767_cast_fp16")];
+            tensor<int32, [3]> var_3460_begin_0 = const()[name = string("op_3460_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_3460_end_0 = const()[name = string("op_3460_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3460_end_mask_0 = const()[name = string("op_3460_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_3460_cast_fp16 = slice_by_index(begin = var_3460_begin_0, end = var_3460_end_0, end_mask = var_3460_end_mask_0, x = cache_57_cast_fp16)[name = string("op_3460_cast_fp16")];
+            bool var_3466_interleave_0 = const()[name = string("op_3466_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3466_cast_fp16 = concat(axis = var_68, interleave = var_3466_interleave_0, values = (var_3460_cast_fp16, key_29_cast_fp16))[name = string("op_3466_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291094784))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291881280))))[name = string("encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291881472)))];
+            tensor<fp16, [1, 7, 1024]> linear_129_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_q_weight_to_fp16_palettized, x = key_29_cast_fp16)[name = string("linear_129_cast_fp16")];
+            tensor<int32, [4]> var_3471 = const()[name = string("op_3471"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_85_cast_fp16 = reshape(shape = var_3471, x = linear_129_cast_fp16)[name = string("q_85_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(291883584))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292670080))))[name = string("encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292670272)))];
+            tensor<fp16, [1, 49, 1024]> linear_130_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_k_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_130_cast_fp16")];
+            tensor<int32, [4]> var_3476 = const()[name = string("op_3476"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_57_cast_fp16 = reshape(shape = var_3476, x = linear_130_cast_fp16)[name = string("k_57_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292672384))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293458880))))[name = string("encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293459072)))];
+            tensor<fp16, [1, 49, 1024]> linear_131_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_v_weight_to_fp16_palettized, x = input_767_cast_fp16)[name = string("linear_131_cast_fp16")];
+            tensor<int32, [4]> var_3481 = const()[name = string("op_3481"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_29_cast_fp16 = reshape(shape = var_3481, x = linear_131_cast_fp16)[name = string("v_29_cast_fp16")];
+            tensor<int32, [4]> value_37_perm_0 = const()[name = string("value_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293461184)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3494_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_u_to_fp16)[name = string("op_3494_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293463296)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3496_cast_fp16 = add(x = q_85_cast_fp16, y = encoder_layers_14_self_attn_pos_bias_v_to_fp16)[name = string("op_3496_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_29_perm_0 = const()[name = string("q_with_bias_v_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_371_transpose_x_0 = const()[name = string("x_371_transpose_x_0"), val = bool(false)];
+            bool x_371_transpose_y_0 = const()[name = string("x_371_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_3498_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293465408))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293564800))))[name = string("op_3498_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3496_cast_fp16)[name = string("transpose_236")];
+            tensor<fp16, [1, 8, 7, 97]> x_371_cast_fp16 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = op_3498_to_fp16_quantized)[name = string("x_371_cast_fp16")];
+            tensor<int32, [8]> x_373_pad_0 = const()[name = string("x_373_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_373_mode_0 = const()[name = string("x_373_mode_0"), val = string("constant")];
+            fp16 const_261_to_fp16 = const()[name = string("const_261_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_373_cast_fp16 = pad(constant_val = const_261_to_fp16, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371_cast_fp16)[name = string("x_373_cast_fp16")];
+            tensor<int32, [4]> var_3506 = const()[name = string("op_3506"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_375_cast_fp16 = reshape(shape = var_3506, x = x_373_cast_fp16)[name = string("x_375_cast_fp16")];
+            tensor<int32, [4]> var_3510_begin_0 = const()[name = string("op_3510_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3510_end_0 = const()[name = string("op_3510_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_3510_end_mask_0 = const()[name = string("op_3510_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_3510_cast_fp16 = slice_by_index(begin = var_3510_begin_0, end = var_3510_end_0, end_mask = var_3510_end_mask_0, x = x_375_cast_fp16)[name = string("op_3510_cast_fp16")];
+            tensor<int32, [4]> var_3511 = const()[name = string("op_3511"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_57_cast_fp16 = reshape(shape = var_3511, x = var_3510_cast_fp16)[name = string("matrix_bd_57_cast_fp16")];
+            bool matrix_ac_29_transpose_x_0 = const()[name = string("matrix_ac_29_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_29_transpose_y_0 = const()[name = string("matrix_ac_29_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_124_perm_0 = const()[name = string("transpose_124_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_125_perm_0 = const()[name = string("transpose_125_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_125 = transpose(perm = transpose_125_perm_0, x = k_57_cast_fp16)[name = string("transpose_234")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_124 = transpose(perm = transpose_124_perm_0, x = var_3494_cast_fp16)[name = string("transpose_235")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_29_cast_fp16 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_124, y = transpose_125)[name = string("matrix_ac_29_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_59_begin_0 = const()[name = string("matrix_bd_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_59_end_0 = const()[name = string("matrix_bd_59_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_59_end_mask_0 = const()[name = string("matrix_bd_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_59_cast_fp16 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57_cast_fp16)[name = string("matrix_bd_59_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3520_cast_fp16 = add(x = matrix_ac_29_cast_fp16, y = matrix_bd_59_cast_fp16)[name = string("op_3520_cast_fp16")];
+            fp16 _inversed_scores_57_y_0_to_fp16 = const()[name = string("_inversed_scores_57_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_57_cast_fp16 = mul(x = var_3520_cast_fp16, y = _inversed_scores_57_y_0_to_fp16)[name = string("_inversed_scores_57_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_59_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_57_cast_fp16, cond = mask_11)[name = string("scores_59_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3526_cast_fp16 = softmax(axis = var_59, x = scores_59_cast_fp16)[name = string("op_3526_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_769_cast_fp16 = select(a = var_44_to_fp16, b = var_3526_cast_fp16, cond = mask_11)[name = string("input_769_cast_fp16")];
+            bool x_377_transpose_x_0 = const()[name = string("x_377_transpose_x_0"), val = bool(false)];
+            bool x_377_transpose_y_0 = const()[name = string("x_377_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_37_cast_fp16 = transpose(perm = value_37_perm_0, x = v_29_cast_fp16)[name = string("transpose_233")];
+            tensor<fp16, [1, 8, 7, 128]> x_377_cast_fp16 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769_cast_fp16, y = value_37_cast_fp16)[name = string("x_377_cast_fp16")];
+            tensor<int32, [4]> var_3530_perm_0 = const()[name = string("op_3530_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3531 = const()[name = string("op_3531"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_3530_cast_fp16 = transpose(perm = var_3530_perm_0, x = x_377_cast_fp16)[name = string("transpose_232")];
+            tensor<fp16, [1, 7, 1024]> input_771_cast_fp16 = reshape(shape = var_3531, x = var_3530_cast_fp16)[name = string("input_771_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293565120))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294351616))))[name = string("encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_14_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294351808)))];
+            tensor<fp16, [1, 7, 1024]> linear_133_cast_fp16 = linear(bias = encoder_layers_14_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_14_self_attn_linear_out_weight_to_fp16_palettized, x = input_771_cast_fp16)[name = string("linear_133_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_775_cast_fp16 = add(x = input_765_cast_fp16, y = linear_133_cast_fp16)[name = string("input_775_cast_fp16")];
+            tensor<int32, [1]> x_381_axes_0 = const()[name = string("x_381_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294353920)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294356032)))];
+            tensor<fp16, [1, 7, 1024]> x_381_cast_fp16 = layer_norm(axes = x_381_axes_0, beta = encoder_layers_14_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_conv_weight_to_fp16, x = input_775_cast_fp16)[name = string("x_381_cast_fp16")];
+            tensor<int32, [3]> input_777_perm_0 = const()[name = string("input_777_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_779_pad_type_0 = const()[name = string("input_779_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_779_strides_0 = const()[name = string("input_779_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_779_pad_0 = const()[name = string("input_779_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_779_dilations_0 = const()[name = string("input_779_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_779_groups_0 = const()[name = string("input_779_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294358144))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296455360))))[name = string("encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_777_cast_fp16 = transpose(perm = input_777_perm_0, x = x_381_cast_fp16)[name = string("transpose_231")];
+            tensor<fp16, [1, 2048, 7]> input_779_cast_fp16 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = encoder_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_777_cast_fp16)[name = string("input_779_cast_fp16")];
+            int32 x_383_split_num_splits_0 = const()[name = string("x_383_split_num_splits_0"), val = int32(2)];
+            int32 x_383_split_axis_0 = const()[name = string("x_383_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_383_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_383_split_cast_fp16_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779_cast_fp16)[name = string("x_383_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_383_split_1_sigmoid_cast_fp16 = sigmoid(x = x_383_split_cast_fp16_1)[name = string("x_383_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_383_cast_fp16 = mul(x = x_383_split_cast_fp16_0, y = x_383_split_1_sigmoid_cast_fp16)[name = string("x_383_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_781_cast_fp16 = select(a = var_44_to_fp16, b = x_383_cast_fp16, cond = var_575)[name = string("input_781_cast_fp16")];
+            bool new_x_59_interleave_0 = const()[name = string("new_x_59_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_59_cast_fp16 = concat(axis = var_59, interleave = new_x_59_interleave_0, values = (cache_59_cast_fp16, input_781_cast_fp16))[name = string("new_x_59_cast_fp16")];
+            tensor<int32, [3]> var_3570_begin_0 = const()[name = string("op_3570_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_3570_end_0 = const()[name = string("op_3570_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_3570_end_mask_0 = const()[name = string("op_3570_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3570_cast_fp16 = slice_by_index(begin = var_3570_begin_0, end = var_3570_end_0, end_mask = var_3570_end_mask_0, x = new_x_59_cast_fp16)[name = string("op_3570_cast_fp16")];
+            string x_385_pad_type_0 = const()[name = string("x_385_pad_type_0"), val = string("valid")];
+            int32 x_385_groups_0 = const()[name = string("x_385_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_385_strides_0 = const()[name = string("x_385_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_385_pad_0 = const()[name = string("x_385_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_385_dilations_0 = const()[name = string("x_385_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296459520))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296468800))))[name = string("encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_385_cast_fp16 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = encoder_layers_14_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_59_cast_fp16)[name = string("x_385_cast_fp16")];
+            tensor<int32, [3]> input_783_perm_0 = const()[name = string("input_783_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_387_axes_0 = const()[name = string("x_387_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296470912)))];
+            tensor<fp16, [1024]> encoder_layers_14_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_14_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296473024)))];
+            tensor<fp16, [1, 7, 1024]> input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_385_cast_fp16)[name = string("transpose_230")];
+            tensor<fp16, [1, 7, 1024]> x_387_cast_fp16 = layer_norm(axes = x_387_axes_0, beta = encoder_layers_14_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_conv_batch_norm_weight_to_fp16, x = input_783_cast_fp16)[name = string("x_387_cast_fp16")];
+            tensor<int32, [3]> input_785_perm_0 = const()[name = string("input_785_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_785_cast_fp16 = transpose(perm = input_785_perm_0, x = x_387_cast_fp16)[name = string("transpose_229")];
+            tensor<fp16, [1, 1024, 7]> input_787_cast_fp16 = silu(x = input_785_cast_fp16)[name = string("input_787_cast_fp16")];
+            string x_389_pad_type_0 = const()[name = string("x_389_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_389_strides_0 = const()[name = string("x_389_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_389_pad_0 = const()[name = string("x_389_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_389_dilations_0 = const()[name = string("x_389_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_389_groups_0 = const()[name = string("x_389_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296475136))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297523776))))[name = string("encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_389_cast_fp16 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = encoder_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_787_cast_fp16)[name = string("x_389_cast_fp16")];
+            tensor<int32, [3]> input_789_perm_0 = const()[name = string("input_789_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_789_cast_fp16 = transpose(perm = input_789_perm_0, x = x_389_cast_fp16)[name = string("transpose_228")];
+            tensor<fp16, [1, 7, 1024]> input_791_cast_fp16 = add(x = input_775_cast_fp16, y = input_789_cast_fp16)[name = string("input_791_cast_fp16")];
+            tensor<int32, [1]> input_793_axes_0 = const()[name = string("input_793_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297525888)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297528000)))];
+            tensor<fp16, [1, 7, 1024]> input_793_cast_fp16 = layer_norm(axes = input_793_axes_0, beta = encoder_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_feed_forward2_weight_to_fp16, x = input_791_cast_fp16)[name = string("input_793_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297530112))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300675904))))[name = string("encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_14_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300676096)))];
+            tensor<fp16, [1, 7, 4096]> linear_134_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear1_weight_to_fp16_palettized, x = input_793_cast_fp16)[name = string("linear_134_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_797_cast_fp16 = silu(x = linear_134_cast_fp16)[name = string("input_797_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300684352))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303830144))))[name = string("encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_14_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_14_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303830336)))];
+            tensor<fp16, [1, 7, 1024]> linear_135_cast_fp16 = linear(bias = encoder_layers_14_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_14_feed_forward2_linear2_weight_to_fp16_palettized, x = input_797_cast_fp16)[name = string("linear_135_cast_fp16")];
+            fp16 var_3613_to_fp16 = const()[name = string("op_3613_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_3614_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_3613_to_fp16)[name = string("op_3614_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_803_cast_fp16 = add(x = input_791_cast_fp16, y = var_3614_cast_fp16)[name = string("input_803_cast_fp16")];
+            tensor<int32, [1]> input_805_axes_0 = const()[name = string("input_805_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_14_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_14_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303832448)))];
+            tensor<fp16, [1024]> encoder_layers_14_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_14_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303834560)))];
+            tensor<fp16, [1, 7, 1024]> input_805_cast_fp16 = layer_norm(axes = input_805_axes_0, beta = encoder_layers_14_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_14_norm_out_weight_to_fp16, x = input_803_cast_fp16)[name = string("input_805_cast_fp16")];
+            tensor<int32, [4]> cache_61_begin_0 = const()[name = string("cache_61_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_61_end_0 = const()[name = string("cache_61_end_0"), val = tensor<int32, [4]>([16, 1, 42, 1024])];
+            tensor<bool, [4]> cache_61_end_mask_0 = const()[name = string("cache_61_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_61_squeeze_mask_0 = const()[name = string("cache_61_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_61_cast_fp16 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_61_cast_fp16")];
+            tensor<int32, [4]> cache_63_begin_0 = const()[name = string("cache_63_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_63_end_0 = const()[name = string("cache_63_end_0"), val = tensor<int32, [4]>([16, 1, 1024, 8])];
+            tensor<bool, [4]> cache_63_end_mask_0 = const()[name = string("cache_63_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_63_squeeze_mask_0 = const()[name = string("cache_63_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_63_cast_fp16 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_63_cast_fp16")];
+            tensor<int32, [1]> input_807_axes_0 = const()[name = string("input_807_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303836672)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303838784)))];
+            tensor<fp16, [1, 7, 1024]> input_807_cast_fp16 = layer_norm(axes = input_807_axes_0, beta = encoder_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_feed_forward1_weight_to_fp16, x = input_805_cast_fp16)[name = string("input_807_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(303840896))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306986688))))[name = string("encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_15_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306986880)))];
+            tensor<fp16, [1, 7, 4096]> linear_136_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear1_weight_to_fp16_palettized, x = input_807_cast_fp16)[name = string("linear_136_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_811_cast_fp16 = silu(x = linear_136_cast_fp16)[name = string("input_811_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306995136))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310140928))))[name = string("encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310141120)))];
+            tensor<fp16, [1, 7, 1024]> linear_137_cast_fp16 = linear(bias = encoder_layers_15_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward1_linear2_weight_to_fp16_palettized, x = input_811_cast_fp16)[name = string("linear_137_cast_fp16")];
+            fp16 var_3650_to_fp16 = const()[name = string("op_3650_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_3651_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_3650_to_fp16)[name = string("op_3651_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_817_cast_fp16 = add(x = input_805_cast_fp16, y = var_3651_cast_fp16)[name = string("input_817_cast_fp16")];
+            tensor<int32, [1]> key_31_axes_0 = const()[name = string("key_31_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310143232)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310145344)))];
+            tensor<fp16, [1, 7, 1024]> key_31_cast_fp16 = layer_norm(axes = key_31_axes_0, beta = encoder_layers_15_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_self_att_weight_to_fp16, x = input_817_cast_fp16)[name = string("key_31_cast_fp16")];
+            bool input_819_interleave_0 = const()[name = string("input_819_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_819_cast_fp16 = concat(axis = var_68, interleave = input_819_interleave_0, values = (cache_61_cast_fp16, key_31_cast_fp16))[name = string("input_819_cast_fp16")];
+            tensor<int32, [3]> var_3673_begin_0 = const()[name = string("op_3673_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_3673_end_0 = const()[name = string("op_3673_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3673_end_mask_0 = const()[name = string("op_3673_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_3673_cast_fp16 = slice_by_index(begin = var_3673_begin_0, end = var_3673_end_0, end_mask = var_3673_end_mask_0, x = cache_61_cast_fp16)[name = string("op_3673_cast_fp16")];
+            bool var_3679_interleave_0 = const()[name = string("op_3679_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3679_cast_fp16 = concat(axis = var_68, interleave = var_3679_interleave_0, values = (var_3673_cast_fp16, key_31_cast_fp16))[name = string("op_3679_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310147456))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310933952))))[name = string("encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310934144)))];
+            tensor<fp16, [1, 7, 1024]> linear_138_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_q_weight_to_fp16_palettized, x = key_31_cast_fp16)[name = string("linear_138_cast_fp16")];
+            tensor<int32, [4]> var_3684 = const()[name = string("op_3684"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_91_cast_fp16 = reshape(shape = var_3684, x = linear_138_cast_fp16)[name = string("q_91_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310936256))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311722752))))[name = string("encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311722944)))];
+            tensor<fp16, [1, 49, 1024]> linear_139_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_k_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_139_cast_fp16")];
+            tensor<int32, [4]> var_3689 = const()[name = string("op_3689"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_61_cast_fp16 = reshape(shape = var_3689, x = linear_139_cast_fp16)[name = string("k_61_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311725056))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312511552))))[name = string("encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312511744)))];
+            tensor<fp16, [1, 49, 1024]> linear_140_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_v_weight_to_fp16_palettized, x = input_819_cast_fp16)[name = string("linear_140_cast_fp16")];
+            tensor<int32, [4]> var_3694 = const()[name = string("op_3694"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_31_cast_fp16 = reshape(shape = var_3694, x = linear_140_cast_fp16)[name = string("v_31_cast_fp16")];
+            tensor<int32, [4]> value_39_perm_0 = const()[name = string("value_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312513856)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3707_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_u_to_fp16)[name = string("op_3707_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312515968)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3709_cast_fp16 = add(x = q_91_cast_fp16, y = encoder_layers_15_self_attn_pos_bias_v_to_fp16)[name = string("op_3709_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_31_perm_0 = const()[name = string("q_with_bias_v_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_397_transpose_x_0 = const()[name = string("x_397_transpose_x_0"), val = bool(false)];
+            bool x_397_transpose_y_0 = const()[name = string("x_397_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_3711_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312518080))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312617472))))[name = string("op_3711_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3709_cast_fp16)[name = string("transpose_227")];
+            tensor<fp16, [1, 8, 7, 97]> x_397_cast_fp16 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = op_3711_to_fp16_quantized)[name = string("x_397_cast_fp16")];
+            tensor<int32, [8]> x_399_pad_0 = const()[name = string("x_399_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_399_mode_0 = const()[name = string("x_399_mode_0"), val = string("constant")];
+            fp16 const_274_to_fp16 = const()[name = string("const_274_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_399_cast_fp16 = pad(constant_val = const_274_to_fp16, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397_cast_fp16)[name = string("x_399_cast_fp16")];
+            tensor<int32, [4]> var_3719 = const()[name = string("op_3719"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_401_cast_fp16 = reshape(shape = var_3719, x = x_399_cast_fp16)[name = string("x_401_cast_fp16")];
+            tensor<int32, [4]> var_3723_begin_0 = const()[name = string("op_3723_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3723_end_0 = const()[name = string("op_3723_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_3723_end_mask_0 = const()[name = string("op_3723_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_3723_cast_fp16 = slice_by_index(begin = var_3723_begin_0, end = var_3723_end_0, end_mask = var_3723_end_mask_0, x = x_401_cast_fp16)[name = string("op_3723_cast_fp16")];
+            tensor<int32, [4]> var_3724 = const()[name = string("op_3724"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_61_cast_fp16 = reshape(shape = var_3724, x = var_3723_cast_fp16)[name = string("matrix_bd_61_cast_fp16")];
+            bool matrix_ac_31_transpose_x_0 = const()[name = string("matrix_ac_31_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_31_transpose_y_0 = const()[name = string("matrix_ac_31_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_126_perm_0 = const()[name = string("transpose_126_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_127_perm_0 = const()[name = string("transpose_127_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_127 = transpose(perm = transpose_127_perm_0, x = k_61_cast_fp16)[name = string("transpose_225")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_126 = transpose(perm = transpose_126_perm_0, x = var_3707_cast_fp16)[name = string("transpose_226")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_31_cast_fp16 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_126, y = transpose_127)[name = string("matrix_ac_31_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_63_begin_0 = const()[name = string("matrix_bd_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_63_end_0 = const()[name = string("matrix_bd_63_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_63_end_mask_0 = const()[name = string("matrix_bd_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_63_cast_fp16 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61_cast_fp16)[name = string("matrix_bd_63_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3733_cast_fp16 = add(x = matrix_ac_31_cast_fp16, y = matrix_bd_63_cast_fp16)[name = string("op_3733_cast_fp16")];
+            fp16 _inversed_scores_61_y_0_to_fp16 = const()[name = string("_inversed_scores_61_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_61_cast_fp16 = mul(x = var_3733_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = string("_inversed_scores_61_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_63_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_61_cast_fp16, cond = mask_11)[name = string("scores_63_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3739_cast_fp16 = softmax(axis = var_59, x = scores_63_cast_fp16)[name = string("op_3739_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_821_cast_fp16 = select(a = var_44_to_fp16, b = var_3739_cast_fp16, cond = mask_11)[name = string("input_821_cast_fp16")];
+            bool x_403_transpose_x_0 = const()[name = string("x_403_transpose_x_0"), val = bool(false)];
+            bool x_403_transpose_y_0 = const()[name = string("x_403_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_39_cast_fp16 = transpose(perm = value_39_perm_0, x = v_31_cast_fp16)[name = string("transpose_224")];
+            tensor<fp16, [1, 8, 7, 128]> x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821_cast_fp16, y = value_39_cast_fp16)[name = string("x_403_cast_fp16")];
+            tensor<int32, [4]> var_3743_perm_0 = const()[name = string("op_3743_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3744 = const()[name = string("op_3744"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_3743_cast_fp16 = transpose(perm = var_3743_perm_0, x = x_403_cast_fp16)[name = string("transpose_223")];
+            tensor<fp16, [1, 7, 1024]> input_823_cast_fp16 = reshape(shape = var_3744, x = var_3743_cast_fp16)[name = string("input_823_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(312617792))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313404288))))[name = string("encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_15_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313404480)))];
+            tensor<fp16, [1, 7, 1024]> linear_142_cast_fp16 = linear(bias = encoder_layers_15_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_15_self_attn_linear_out_weight_to_fp16_palettized, x = input_823_cast_fp16)[name = string("linear_142_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_827_cast_fp16 = add(x = input_817_cast_fp16, y = linear_142_cast_fp16)[name = string("input_827_cast_fp16")];
+            tensor<int32, [1]> x_407_axes_0 = const()[name = string("x_407_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313406592)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313408704)))];
+            tensor<fp16, [1, 7, 1024]> x_407_cast_fp16 = layer_norm(axes = x_407_axes_0, beta = encoder_layers_15_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_conv_weight_to_fp16, x = input_827_cast_fp16)[name = string("x_407_cast_fp16")];
+            tensor<int32, [3]> input_829_perm_0 = const()[name = string("input_829_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_831_pad_type_0 = const()[name = string("input_831_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_831_strides_0 = const()[name = string("input_831_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_831_pad_0 = const()[name = string("input_831_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_831_dilations_0 = const()[name = string("input_831_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_831_groups_0 = const()[name = string("input_831_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313410816))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315508032))))[name = string("encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_829_cast_fp16 = transpose(perm = input_829_perm_0, x = x_407_cast_fp16)[name = string("transpose_222")];
+            tensor<fp16, [1, 2048, 7]> input_831_cast_fp16 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = encoder_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_829_cast_fp16)[name = string("input_831_cast_fp16")];
+            int32 x_409_split_num_splits_0 = const()[name = string("x_409_split_num_splits_0"), val = int32(2)];
+            int32 x_409_split_axis_0 = const()[name = string("x_409_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_409_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_409_split_cast_fp16_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831_cast_fp16)[name = string("x_409_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_409_split_1_sigmoid_cast_fp16 = sigmoid(x = x_409_split_cast_fp16_1)[name = string("x_409_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_409_cast_fp16 = mul(x = x_409_split_cast_fp16_0, y = x_409_split_1_sigmoid_cast_fp16)[name = string("x_409_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_833_cast_fp16 = select(a = var_44_to_fp16, b = x_409_cast_fp16, cond = var_575)[name = string("input_833_cast_fp16")];
+            bool new_x_63_interleave_0 = const()[name = string("new_x_63_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_63_cast_fp16 = concat(axis = var_59, interleave = new_x_63_interleave_0, values = (cache_63_cast_fp16, input_833_cast_fp16))[name = string("new_x_63_cast_fp16")];
+            tensor<int32, [3]> var_3783_begin_0 = const()[name = string("op_3783_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_3783_end_0 = const()[name = string("op_3783_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_3783_end_mask_0 = const()[name = string("op_3783_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3783_cast_fp16 = slice_by_index(begin = var_3783_begin_0, end = var_3783_end_0, end_mask = var_3783_end_mask_0, x = new_x_63_cast_fp16)[name = string("op_3783_cast_fp16")];
+            string x_411_pad_type_0 = const()[name = string("x_411_pad_type_0"), val = string("valid")];
+            int32 x_411_groups_0 = const()[name = string("x_411_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_411_strides_0 = const()[name = string("x_411_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_411_pad_0 = const()[name = string("x_411_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_411_dilations_0 = const()[name = string("x_411_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315512192))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315521472))))[name = string("encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_411_cast_fp16 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = encoder_layers_15_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_63_cast_fp16)[name = string("x_411_cast_fp16")];
+            tensor<int32, [3]> input_835_perm_0 = const()[name = string("input_835_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_413_axes_0 = const()[name = string("x_413_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315523584)))];
+            tensor<fp16, [1024]> encoder_layers_15_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_15_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315525696)))];
+            tensor<fp16, [1, 7, 1024]> input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_411_cast_fp16)[name = string("transpose_221")];
+            tensor<fp16, [1, 7, 1024]> x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = encoder_layers_15_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_conv_batch_norm_weight_to_fp16, x = input_835_cast_fp16)[name = string("x_413_cast_fp16")];
+            tensor<int32, [3]> input_837_perm_0 = const()[name = string("input_837_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_837_cast_fp16 = transpose(perm = input_837_perm_0, x = x_413_cast_fp16)[name = string("transpose_220")];
+            tensor<fp16, [1, 1024, 7]> input_839_cast_fp16 = silu(x = input_837_cast_fp16)[name = string("input_839_cast_fp16")];
+            string x_415_pad_type_0 = const()[name = string("x_415_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_415_strides_0 = const()[name = string("x_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_415_pad_0 = const()[name = string("x_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_415_dilations_0 = const()[name = string("x_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_415_groups_0 = const()[name = string("x_415_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315527808))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316576448))))[name = string("encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_415_cast_fp16 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = encoder_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_839_cast_fp16)[name = string("x_415_cast_fp16")];
+            tensor<int32, [3]> input_841_perm_0 = const()[name = string("input_841_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_841_cast_fp16 = transpose(perm = input_841_perm_0, x = x_415_cast_fp16)[name = string("transpose_219")];
+            tensor<fp16, [1, 7, 1024]> input_843_cast_fp16 = add(x = input_827_cast_fp16, y = input_841_cast_fp16)[name = string("input_843_cast_fp16")];
+            tensor<int32, [1]> input_845_axes_0 = const()[name = string("input_845_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316578560)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316580672)))];
+            tensor<fp16, [1, 7, 1024]> input_845_cast_fp16 = layer_norm(axes = input_845_axes_0, beta = encoder_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_feed_forward2_weight_to_fp16, x = input_843_cast_fp16)[name = string("input_845_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(316582784))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319728576))))[name = string("encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_15_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319728768)))];
+            tensor<fp16, [1, 7, 4096]> linear_143_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear1_weight_to_fp16_palettized, x = input_845_cast_fp16)[name = string("linear_143_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_849_cast_fp16 = silu(x = linear_143_cast_fp16)[name = string("input_849_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319737024))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322882816))))[name = string("encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_15_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_15_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322883008)))];
+            tensor<fp16, [1, 7, 1024]> linear_144_cast_fp16 = linear(bias = encoder_layers_15_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_15_feed_forward2_linear2_weight_to_fp16_palettized, x = input_849_cast_fp16)[name = string("linear_144_cast_fp16")];
+            fp16 var_3826_to_fp16 = const()[name = string("op_3826_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_3827_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_3826_to_fp16)[name = string("op_3827_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_855_cast_fp16 = add(x = input_843_cast_fp16, y = var_3827_cast_fp16)[name = string("input_855_cast_fp16")];
+            tensor<int32, [1]> input_857_axes_0 = const()[name = string("input_857_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_15_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_15_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322885120)))];
+            tensor<fp16, [1024]> encoder_layers_15_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_15_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322887232)))];
+            tensor<fp16, [1, 7, 1024]> input_857_cast_fp16 = layer_norm(axes = input_857_axes_0, beta = encoder_layers_15_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_15_norm_out_weight_to_fp16, x = input_855_cast_fp16)[name = string("input_857_cast_fp16")];
+            tensor<int32, [4]> cache_65_begin_0 = const()[name = string("cache_65_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_65_end_0 = const()[name = string("cache_65_end_0"), val = tensor<int32, [4]>([17, 1, 42, 1024])];
+            tensor<bool, [4]> cache_65_end_mask_0 = const()[name = string("cache_65_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_65_squeeze_mask_0 = const()[name = string("cache_65_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_65_cast_fp16 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_65_cast_fp16")];
+            tensor<int32, [4]> cache_67_begin_0 = const()[name = string("cache_67_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_67_end_0 = const()[name = string("cache_67_end_0"), val = tensor<int32, [4]>([17, 1, 1024, 8])];
+            tensor<bool, [4]> cache_67_end_mask_0 = const()[name = string("cache_67_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_67_squeeze_mask_0 = const()[name = string("cache_67_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_67_cast_fp16 = slice_by_index(begin = cache_67_begin_0, end = cache_67_end_0, end_mask = cache_67_end_mask_0, squeeze_mask = cache_67_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_67_cast_fp16")];
+            tensor<int32, [1]> input_859_axes_0 = const()[name = string("input_859_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322889344)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322891456)))];
+            tensor<fp16, [1, 7, 1024]> input_859_cast_fp16 = layer_norm(axes = input_859_axes_0, beta = encoder_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_feed_forward1_weight_to_fp16, x = input_857_cast_fp16)[name = string("input_859_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322893568))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326039360))))[name = string("encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_16_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326039552)))];
+            tensor<fp16, [1, 7, 4096]> linear_145_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear1_weight_to_fp16_palettized, x = input_859_cast_fp16)[name = string("linear_145_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_863_cast_fp16 = silu(x = linear_145_cast_fp16)[name = string("input_863_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326047808))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329193600))))[name = string("encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329193792)))];
+            tensor<fp16, [1, 7, 1024]> linear_146_cast_fp16 = linear(bias = encoder_layers_16_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward1_linear2_weight_to_fp16_palettized, x = input_863_cast_fp16)[name = string("linear_146_cast_fp16")];
+            fp16 var_3863_to_fp16 = const()[name = string("op_3863_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_3864_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_3863_to_fp16)[name = string("op_3864_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_869_cast_fp16 = add(x = input_857_cast_fp16, y = var_3864_cast_fp16)[name = string("input_869_cast_fp16")];
+            tensor<int32, [1]> key_33_axes_0 = const()[name = string("key_33_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329195904)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329198016)))];
+            tensor<fp16, [1, 7, 1024]> key_33_cast_fp16 = layer_norm(axes = key_33_axes_0, beta = encoder_layers_16_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_self_att_weight_to_fp16, x = input_869_cast_fp16)[name = string("key_33_cast_fp16")];
+            bool input_871_interleave_0 = const()[name = string("input_871_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_871_cast_fp16 = concat(axis = var_68, interleave = input_871_interleave_0, values = (cache_65_cast_fp16, key_33_cast_fp16))[name = string("input_871_cast_fp16")];
+            tensor<int32, [3]> var_3886_begin_0 = const()[name = string("op_3886_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_3886_end_0 = const()[name = string("op_3886_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_3886_end_mask_0 = const()[name = string("op_3886_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_3886_cast_fp16 = slice_by_index(begin = var_3886_begin_0, end = var_3886_end_0, end_mask = var_3886_end_mask_0, x = cache_65_cast_fp16)[name = string("op_3886_cast_fp16")];
+            bool var_3892_interleave_0 = const()[name = string("op_3892_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_3892_cast_fp16 = concat(axis = var_68, interleave = var_3892_interleave_0, values = (var_3886_cast_fp16, key_33_cast_fp16))[name = string("op_3892_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329200128))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329986624))))[name = string("encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329986816)))];
+            tensor<fp16, [1, 7, 1024]> linear_147_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_q_weight_to_fp16_palettized, x = key_33_cast_fp16)[name = string("linear_147_cast_fp16")];
+            tensor<int32, [4]> var_3897 = const()[name = string("op_3897"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_97_cast_fp16 = reshape(shape = var_3897, x = linear_147_cast_fp16)[name = string("q_97_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329988928))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330775424))))[name = string("encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330775616)))];
+            tensor<fp16, [1, 49, 1024]> linear_148_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_k_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_148_cast_fp16")];
+            tensor<int32, [4]> var_3902 = const()[name = string("op_3902"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_65_cast_fp16 = reshape(shape = var_3902, x = linear_148_cast_fp16)[name = string("k_65_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330777728))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331564224))))[name = string("encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331564416)))];
+            tensor<fp16, [1, 49, 1024]> linear_149_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_v_weight_to_fp16_palettized, x = input_871_cast_fp16)[name = string("linear_149_cast_fp16")];
+            tensor<int32, [4]> var_3907 = const()[name = string("op_3907"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_33_cast_fp16 = reshape(shape = var_3907, x = linear_149_cast_fp16)[name = string("v_33_cast_fp16")];
+            tensor<int32, [4]> value_41_perm_0 = const()[name = string("value_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331566528)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3920_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_u_to_fp16)[name = string("op_3920_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331568640)))];
+            tensor<fp16, [1, 7, 8, 128]> var_3922_cast_fp16 = add(x = q_97_cast_fp16, y = encoder_layers_16_self_attn_pos_bias_v_to_fp16)[name = string("op_3922_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_33_perm_0 = const()[name = string("q_with_bias_v_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_423_transpose_x_0 = const()[name = string("x_423_transpose_x_0"), val = bool(false)];
+            bool x_423_transpose_y_0 = const()[name = string("x_423_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_3924_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331570752))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331670144))))[name = string("op_3924_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_33_cast_fp16 = transpose(perm = q_with_bias_v_33_perm_0, x = var_3922_cast_fp16)[name = string("transpose_218")];
+            tensor<fp16, [1, 8, 7, 97]> x_423_cast_fp16 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_33_cast_fp16, y = op_3924_to_fp16_quantized)[name = string("x_423_cast_fp16")];
+            tensor<int32, [8]> x_425_pad_0 = const()[name = string("x_425_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_425_mode_0 = const()[name = string("x_425_mode_0"), val = string("constant")];
+            fp16 const_287_to_fp16 = const()[name = string("const_287_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_425_cast_fp16 = pad(constant_val = const_287_to_fp16, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423_cast_fp16)[name = string("x_425_cast_fp16")];
+            tensor<int32, [4]> var_3932 = const()[name = string("op_3932"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_427_cast_fp16 = reshape(shape = var_3932, x = x_425_cast_fp16)[name = string("x_427_cast_fp16")];
+            tensor<int32, [4]> var_3936_begin_0 = const()[name = string("op_3936_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3936_end_0 = const()[name = string("op_3936_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_3936_end_mask_0 = const()[name = string("op_3936_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_3936_cast_fp16 = slice_by_index(begin = var_3936_begin_0, end = var_3936_end_0, end_mask = var_3936_end_mask_0, x = x_427_cast_fp16)[name = string("op_3936_cast_fp16")];
+            tensor<int32, [4]> var_3937 = const()[name = string("op_3937"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_65_cast_fp16 = reshape(shape = var_3937, x = var_3936_cast_fp16)[name = string("matrix_bd_65_cast_fp16")];
+            bool matrix_ac_33_transpose_x_0 = const()[name = string("matrix_ac_33_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_33_transpose_y_0 = const()[name = string("matrix_ac_33_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_128_perm_0 = const()[name = string("transpose_128_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_129_perm_0 = const()[name = string("transpose_129_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_129 = transpose(perm = transpose_129_perm_0, x = k_65_cast_fp16)[name = string("transpose_216")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_128 = transpose(perm = transpose_128_perm_0, x = var_3920_cast_fp16)[name = string("transpose_217")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_33_cast_fp16 = matmul(transpose_x = matrix_ac_33_transpose_x_0, transpose_y = matrix_ac_33_transpose_y_0, x = transpose_128, y = transpose_129)[name = string("matrix_ac_33_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_67_begin_0 = const()[name = string("matrix_bd_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_67_end_0 = const()[name = string("matrix_bd_67_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_67_end_mask_0 = const()[name = string("matrix_bd_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_67_cast_fp16 = slice_by_index(begin = matrix_bd_67_begin_0, end = matrix_bd_67_end_0, end_mask = matrix_bd_67_end_mask_0, x = matrix_bd_65_cast_fp16)[name = string("matrix_bd_67_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3946_cast_fp16 = add(x = matrix_ac_33_cast_fp16, y = matrix_bd_67_cast_fp16)[name = string("op_3946_cast_fp16")];
+            fp16 _inversed_scores_65_y_0_to_fp16 = const()[name = string("_inversed_scores_65_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_65_cast_fp16 = mul(x = var_3946_cast_fp16, y = _inversed_scores_65_y_0_to_fp16)[name = string("_inversed_scores_65_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_67_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_65_cast_fp16, cond = mask_11)[name = string("scores_67_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_3952_cast_fp16 = softmax(axis = var_59, x = scores_67_cast_fp16)[name = string("op_3952_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_873_cast_fp16 = select(a = var_44_to_fp16, b = var_3952_cast_fp16, cond = mask_11)[name = string("input_873_cast_fp16")];
+            bool x_429_transpose_x_0 = const()[name = string("x_429_transpose_x_0"), val = bool(false)];
+            bool x_429_transpose_y_0 = const()[name = string("x_429_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_41_cast_fp16 = transpose(perm = value_41_perm_0, x = v_33_cast_fp16)[name = string("transpose_215")];
+            tensor<fp16, [1, 8, 7, 128]> x_429_cast_fp16 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873_cast_fp16, y = value_41_cast_fp16)[name = string("x_429_cast_fp16")];
+            tensor<int32, [4]> var_3956_perm_0 = const()[name = string("op_3956_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3957 = const()[name = string("op_3957"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_3956_cast_fp16 = transpose(perm = var_3956_perm_0, x = x_429_cast_fp16)[name = string("transpose_214")];
+            tensor<fp16, [1, 7, 1024]> input_875_cast_fp16 = reshape(shape = var_3957, x = var_3956_cast_fp16)[name = string("input_875_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331670464))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332456960))))[name = string("encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_16_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332457152)))];
+            tensor<fp16, [1, 7, 1024]> linear_151_cast_fp16 = linear(bias = encoder_layers_16_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_16_self_attn_linear_out_weight_to_fp16_palettized, x = input_875_cast_fp16)[name = string("linear_151_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_879_cast_fp16 = add(x = input_869_cast_fp16, y = linear_151_cast_fp16)[name = string("input_879_cast_fp16")];
+            tensor<int32, [1]> x_433_axes_0 = const()[name = string("x_433_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332459264)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332461376)))];
+            tensor<fp16, [1, 7, 1024]> x_433_cast_fp16 = layer_norm(axes = x_433_axes_0, beta = encoder_layers_16_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_conv_weight_to_fp16, x = input_879_cast_fp16)[name = string("x_433_cast_fp16")];
+            tensor<int32, [3]> input_881_perm_0 = const()[name = string("input_881_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_883_pad_type_0 = const()[name = string("input_883_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_883_strides_0 = const()[name = string("input_883_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_883_pad_0 = const()[name = string("input_883_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_883_dilations_0 = const()[name = string("input_883_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_883_groups_0 = const()[name = string("input_883_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(332463488))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334560704))))[name = string("encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_881_cast_fp16 = transpose(perm = input_881_perm_0, x = x_433_cast_fp16)[name = string("transpose_213")];
+            tensor<fp16, [1, 2048, 7]> input_883_cast_fp16 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = encoder_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_881_cast_fp16)[name = string("input_883_cast_fp16")];
+            int32 x_435_split_num_splits_0 = const()[name = string("x_435_split_num_splits_0"), val = int32(2)];
+            int32 x_435_split_axis_0 = const()[name = string("x_435_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_435_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_435_split_cast_fp16_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883_cast_fp16)[name = string("x_435_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_435_split_1_sigmoid_cast_fp16 = sigmoid(x = x_435_split_cast_fp16_1)[name = string("x_435_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_435_cast_fp16 = mul(x = x_435_split_cast_fp16_0, y = x_435_split_1_sigmoid_cast_fp16)[name = string("x_435_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_885_cast_fp16 = select(a = var_44_to_fp16, b = x_435_cast_fp16, cond = var_575)[name = string("input_885_cast_fp16")];
+            bool new_x_67_interleave_0 = const()[name = string("new_x_67_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_67_cast_fp16 = concat(axis = var_59, interleave = new_x_67_interleave_0, values = (cache_67_cast_fp16, input_885_cast_fp16))[name = string("new_x_67_cast_fp16")];
+            tensor<int32, [3]> var_3996_begin_0 = const()[name = string("op_3996_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_3996_end_0 = const()[name = string("op_3996_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_3996_end_mask_0 = const()[name = string("op_3996_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_3996_cast_fp16 = slice_by_index(begin = var_3996_begin_0, end = var_3996_end_0, end_mask = var_3996_end_mask_0, x = new_x_67_cast_fp16)[name = string("op_3996_cast_fp16")];
+            string x_437_pad_type_0 = const()[name = string("x_437_pad_type_0"), val = string("valid")];
+            int32 x_437_groups_0 = const()[name = string("x_437_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_437_strides_0 = const()[name = string("x_437_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_437_pad_0 = const()[name = string("x_437_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_437_dilations_0 = const()[name = string("x_437_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334564864))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334574144))))[name = string("encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_437_cast_fp16 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = encoder_layers_16_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_67_cast_fp16)[name = string("x_437_cast_fp16")];
+            tensor<int32, [3]> input_887_perm_0 = const()[name = string("input_887_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_439_axes_0 = const()[name = string("x_439_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334576256)))];
+            tensor<fp16, [1024]> encoder_layers_16_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_16_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334578368)))];
+            tensor<fp16, [1, 7, 1024]> input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_437_cast_fp16)[name = string("transpose_212")];
+            tensor<fp16, [1, 7, 1024]> x_439_cast_fp16 = layer_norm(axes = x_439_axes_0, beta = encoder_layers_16_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_conv_batch_norm_weight_to_fp16, x = input_887_cast_fp16)[name = string("x_439_cast_fp16")];
+            tensor<int32, [3]> input_889_perm_0 = const()[name = string("input_889_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_889_cast_fp16 = transpose(perm = input_889_perm_0, x = x_439_cast_fp16)[name = string("transpose_211")];
+            tensor<fp16, [1, 1024, 7]> input_891_cast_fp16 = silu(x = input_889_cast_fp16)[name = string("input_891_cast_fp16")];
+            string x_441_pad_type_0 = const()[name = string("x_441_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_441_strides_0 = const()[name = string("x_441_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_441_pad_0 = const()[name = string("x_441_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_441_dilations_0 = const()[name = string("x_441_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_441_groups_0 = const()[name = string("x_441_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(334580480))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335629120))))[name = string("encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_441_cast_fp16 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = encoder_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_891_cast_fp16)[name = string("x_441_cast_fp16")];
+            tensor<int32, [3]> input_893_perm_0 = const()[name = string("input_893_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_893_cast_fp16 = transpose(perm = input_893_perm_0, x = x_441_cast_fp16)[name = string("transpose_210")];
+            tensor<fp16, [1, 7, 1024]> input_895_cast_fp16 = add(x = input_879_cast_fp16, y = input_893_cast_fp16)[name = string("input_895_cast_fp16")];
+            tensor<int32, [1]> input_897_axes_0 = const()[name = string("input_897_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335631232)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335633344)))];
+            tensor<fp16, [1, 7, 1024]> input_897_cast_fp16 = layer_norm(axes = input_897_axes_0, beta = encoder_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_feed_forward2_weight_to_fp16, x = input_895_cast_fp16)[name = string("input_897_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335635456))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338781248))))[name = string("encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_16_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338781440)))];
+            tensor<fp16, [1, 7, 4096]> linear_152_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear1_weight_to_fp16_palettized, x = input_897_cast_fp16)[name = string("linear_152_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_901_cast_fp16 = silu(x = linear_152_cast_fp16)[name = string("input_901_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(338789696))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341935488))))[name = string("encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_16_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_16_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341935680)))];
+            tensor<fp16, [1, 7, 1024]> linear_153_cast_fp16 = linear(bias = encoder_layers_16_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_16_feed_forward2_linear2_weight_to_fp16_palettized, x = input_901_cast_fp16)[name = string("linear_153_cast_fp16")];
+            fp16 var_4039_to_fp16 = const()[name = string("op_4039_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4040_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_4039_to_fp16)[name = string("op_4040_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_907_cast_fp16 = add(x = input_895_cast_fp16, y = var_4040_cast_fp16)[name = string("input_907_cast_fp16")];
+            tensor<int32, [1]> input_909_axes_0 = const()[name = string("input_909_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_16_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_16_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341937792)))];
+            tensor<fp16, [1024]> encoder_layers_16_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_16_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341939904)))];
+            tensor<fp16, [1, 7, 1024]> input_909_cast_fp16 = layer_norm(axes = input_909_axes_0, beta = encoder_layers_16_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_16_norm_out_weight_to_fp16, x = input_907_cast_fp16)[name = string("input_909_cast_fp16")];
+            tensor<int32, [4]> cache_69_begin_0 = const()[name = string("cache_69_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> cache_69_end_0 = const()[name = string("cache_69_end_0"), val = tensor<int32, [4]>([18, 1, 42, 1024])];
+            tensor<bool, [4]> cache_69_end_mask_0 = const()[name = string("cache_69_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_69_squeeze_mask_0 = const()[name = string("cache_69_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_69_cast_fp16 = slice_by_index(begin = cache_69_begin_0, end = cache_69_end_0, end_mask = cache_69_end_mask_0, squeeze_mask = cache_69_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_69_cast_fp16")];
+            tensor<int32, [4]> cache_71_begin_0 = const()[name = string("cache_71_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> cache_71_end_0 = const()[name = string("cache_71_end_0"), val = tensor<int32, [4]>([18, 1, 1024, 8])];
+            tensor<bool, [4]> cache_71_end_mask_0 = const()[name = string("cache_71_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_71_squeeze_mask_0 = const()[name = string("cache_71_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_71_cast_fp16 = slice_by_index(begin = cache_71_begin_0, end = cache_71_end_0, end_mask = cache_71_end_mask_0, squeeze_mask = cache_71_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_71_cast_fp16")];
+            tensor<int32, [1]> input_911_axes_0 = const()[name = string("input_911_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341942016)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341944128)))];
+            tensor<fp16, [1, 7, 1024]> input_911_cast_fp16 = layer_norm(axes = input_911_axes_0, beta = encoder_layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_feed_forward1_weight_to_fp16, x = input_909_cast_fp16)[name = string("input_911_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(341946240))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345092032))))[name = string("encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_17_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345092224)))];
+            tensor<fp16, [1, 7, 4096]> linear_154_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear1_weight_to_fp16_palettized, x = input_911_cast_fp16)[name = string("linear_154_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_915_cast_fp16 = silu(x = linear_154_cast_fp16)[name = string("input_915_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345100480))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348246272))))[name = string("encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348246464)))];
+            tensor<fp16, [1, 7, 1024]> linear_155_cast_fp16 = linear(bias = encoder_layers_17_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward1_linear2_weight_to_fp16_palettized, x = input_915_cast_fp16)[name = string("linear_155_cast_fp16")];
+            fp16 var_4076_to_fp16 = const()[name = string("op_4076_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4077_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_4076_to_fp16)[name = string("op_4077_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_921_cast_fp16 = add(x = input_909_cast_fp16, y = var_4077_cast_fp16)[name = string("input_921_cast_fp16")];
+            tensor<int32, [1]> key_35_axes_0 = const()[name = string("key_35_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348248576)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348250688)))];
+            tensor<fp16, [1, 7, 1024]> key_35_cast_fp16 = layer_norm(axes = key_35_axes_0, beta = encoder_layers_17_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_self_att_weight_to_fp16, x = input_921_cast_fp16)[name = string("key_35_cast_fp16")];
+            bool input_923_interleave_0 = const()[name = string("input_923_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_923_cast_fp16 = concat(axis = var_68, interleave = input_923_interleave_0, values = (cache_69_cast_fp16, key_35_cast_fp16))[name = string("input_923_cast_fp16")];
+            tensor<int32, [3]> var_4099_begin_0 = const()[name = string("op_4099_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_4099_end_0 = const()[name = string("op_4099_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4099_end_mask_0 = const()[name = string("op_4099_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_4099_cast_fp16 = slice_by_index(begin = var_4099_begin_0, end = var_4099_end_0, end_mask = var_4099_end_mask_0, x = cache_69_cast_fp16)[name = string("op_4099_cast_fp16")];
+            bool var_4105_interleave_0 = const()[name = string("op_4105_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4105_cast_fp16 = concat(axis = var_68, interleave = var_4105_interleave_0, values = (var_4099_cast_fp16, key_35_cast_fp16))[name = string("op_4105_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(348252800))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349039296))))[name = string("encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349039488)))];
+            tensor<fp16, [1, 7, 1024]> linear_156_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_q_weight_to_fp16_palettized, x = key_35_cast_fp16)[name = string("linear_156_cast_fp16")];
+            tensor<int32, [4]> var_4110 = const()[name = string("op_4110"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_103_cast_fp16 = reshape(shape = var_4110, x = linear_156_cast_fp16)[name = string("q_103_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349041600))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349828096))))[name = string("encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349828288)))];
+            tensor<fp16, [1, 49, 1024]> linear_157_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_k_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_157_cast_fp16")];
+            tensor<int32, [4]> var_4115 = const()[name = string("op_4115"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_69_cast_fp16 = reshape(shape = var_4115, x = linear_157_cast_fp16)[name = string("k_69_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349830400))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350616896))))[name = string("encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350617088)))];
+            tensor<fp16, [1, 49, 1024]> linear_158_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_v_weight_to_fp16_palettized, x = input_923_cast_fp16)[name = string("linear_158_cast_fp16")];
+            tensor<int32, [4]> var_4120 = const()[name = string("op_4120"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_35_cast_fp16 = reshape(shape = var_4120, x = linear_158_cast_fp16)[name = string("v_35_cast_fp16")];
+            tensor<int32, [4]> value_43_perm_0 = const()[name = string("value_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_17_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350619200)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4133_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_u_to_fp16)[name = string("op_4133_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_17_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_17_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350621312)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4135_cast_fp16 = add(x = q_103_cast_fp16, y = encoder_layers_17_self_attn_pos_bias_v_to_fp16)[name = string("op_4135_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_35_perm_0 = const()[name = string("q_with_bias_v_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_449_transpose_x_0 = const()[name = string("x_449_transpose_x_0"), val = bool(false)];
+            bool x_449_transpose_y_0 = const()[name = string("x_449_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_4137_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350623424))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350722816))))[name = string("op_4137_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_35_cast_fp16 = transpose(perm = q_with_bias_v_35_perm_0, x = var_4135_cast_fp16)[name = string("transpose_209")];
+            tensor<fp16, [1, 8, 7, 97]> x_449_cast_fp16 = matmul(transpose_x = x_449_transpose_x_0, transpose_y = x_449_transpose_y_0, x = q_with_bias_v_35_cast_fp16, y = op_4137_to_fp16_quantized)[name = string("x_449_cast_fp16")];
+            tensor<int32, [8]> x_451_pad_0 = const()[name = string("x_451_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_451_mode_0 = const()[name = string("x_451_mode_0"), val = string("constant")];
+            fp16 const_300_to_fp16 = const()[name = string("const_300_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_451_cast_fp16 = pad(constant_val = const_300_to_fp16, mode = x_451_mode_0, pad = x_451_pad_0, x = x_449_cast_fp16)[name = string("x_451_cast_fp16")];
+            tensor<int32, [4]> var_4145 = const()[name = string("op_4145"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_453_cast_fp16 = reshape(shape = var_4145, x = x_451_cast_fp16)[name = string("x_453_cast_fp16")];
+            tensor<int32, [4]> var_4149_begin_0 = const()[name = string("op_4149_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4149_end_0 = const()[name = string("op_4149_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_4149_end_mask_0 = const()[name = string("op_4149_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_4149_cast_fp16 = slice_by_index(begin = var_4149_begin_0, end = var_4149_end_0, end_mask = var_4149_end_mask_0, x = x_453_cast_fp16)[name = string("op_4149_cast_fp16")];
+            tensor<int32, [4]> var_4150 = const()[name = string("op_4150"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_69_cast_fp16 = reshape(shape = var_4150, x = var_4149_cast_fp16)[name = string("matrix_bd_69_cast_fp16")];
+            bool matrix_ac_35_transpose_x_0 = const()[name = string("matrix_ac_35_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_35_transpose_y_0 = const()[name = string("matrix_ac_35_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_130_perm_0 = const()[name = string("transpose_130_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_131_perm_0 = const()[name = string("transpose_131_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_131 = transpose(perm = transpose_131_perm_0, x = k_69_cast_fp16)[name = string("transpose_207")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_130 = transpose(perm = transpose_130_perm_0, x = var_4133_cast_fp16)[name = string("transpose_208")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_35_cast_fp16 = matmul(transpose_x = matrix_ac_35_transpose_x_0, transpose_y = matrix_ac_35_transpose_y_0, x = transpose_130, y = transpose_131)[name = string("matrix_ac_35_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_71_begin_0 = const()[name = string("matrix_bd_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_71_end_0 = const()[name = string("matrix_bd_71_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_71_end_mask_0 = const()[name = string("matrix_bd_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_71_cast_fp16 = slice_by_index(begin = matrix_bd_71_begin_0, end = matrix_bd_71_end_0, end_mask = matrix_bd_71_end_mask_0, x = matrix_bd_69_cast_fp16)[name = string("matrix_bd_71_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_4159_cast_fp16 = add(x = matrix_ac_35_cast_fp16, y = matrix_bd_71_cast_fp16)[name = string("op_4159_cast_fp16")];
+            fp16 _inversed_scores_69_y_0_to_fp16 = const()[name = string("_inversed_scores_69_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_69_cast_fp16 = mul(x = var_4159_cast_fp16, y = _inversed_scores_69_y_0_to_fp16)[name = string("_inversed_scores_69_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_71_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_69_cast_fp16, cond = mask_11)[name = string("scores_71_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_4165_cast_fp16 = softmax(axis = var_59, x = scores_71_cast_fp16)[name = string("op_4165_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_925_cast_fp16 = select(a = var_44_to_fp16, b = var_4165_cast_fp16, cond = mask_11)[name = string("input_925_cast_fp16")];
+            bool x_455_transpose_x_0 = const()[name = string("x_455_transpose_x_0"), val = bool(false)];
+            bool x_455_transpose_y_0 = const()[name = string("x_455_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_43_cast_fp16 = transpose(perm = value_43_perm_0, x = v_35_cast_fp16)[name = string("transpose_206")];
+            tensor<fp16, [1, 8, 7, 128]> x_455_cast_fp16 = matmul(transpose_x = x_455_transpose_x_0, transpose_y = x_455_transpose_y_0, x = input_925_cast_fp16, y = value_43_cast_fp16)[name = string("x_455_cast_fp16")];
+            tensor<int32, [4]> var_4169_perm_0 = const()[name = string("op_4169_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4170 = const()[name = string("op_4170"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_4169_cast_fp16 = transpose(perm = var_4169_perm_0, x = x_455_cast_fp16)[name = string("transpose_205")];
+            tensor<fp16, [1, 7, 1024]> input_927_cast_fp16 = reshape(shape = var_4170, x = var_4169_cast_fp16)[name = string("input_927_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(350723136))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351509632))))[name = string("encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_17_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351509824)))];
+            tensor<fp16, [1, 7, 1024]> linear_160_cast_fp16 = linear(bias = encoder_layers_17_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_17_self_attn_linear_out_weight_to_fp16_palettized, x = input_927_cast_fp16)[name = string("linear_160_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_931_cast_fp16 = add(x = input_921_cast_fp16, y = linear_160_cast_fp16)[name = string("input_931_cast_fp16")];
+            tensor<int32, [1]> x_459_axes_0 = const()[name = string("x_459_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351511936)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351514048)))];
+            tensor<fp16, [1, 7, 1024]> x_459_cast_fp16 = layer_norm(axes = x_459_axes_0, beta = encoder_layers_17_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_conv_weight_to_fp16, x = input_931_cast_fp16)[name = string("x_459_cast_fp16")];
+            tensor<int32, [3]> input_933_perm_0 = const()[name = string("input_933_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_935_pad_type_0 = const()[name = string("input_935_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_935_strides_0 = const()[name = string("input_935_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_935_pad_0 = const()[name = string("input_935_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_935_dilations_0 = const()[name = string("input_935_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_935_groups_0 = const()[name = string("input_935_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351516160))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353613376))))[name = string("encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_933_cast_fp16 = transpose(perm = input_933_perm_0, x = x_459_cast_fp16)[name = string("transpose_204")];
+            tensor<fp16, [1, 2048, 7]> input_935_cast_fp16 = conv(dilations = input_935_dilations_0, groups = input_935_groups_0, pad = input_935_pad_0, pad_type = input_935_pad_type_0, strides = input_935_strides_0, weight = encoder_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_933_cast_fp16)[name = string("input_935_cast_fp16")];
+            int32 x_461_split_num_splits_0 = const()[name = string("x_461_split_num_splits_0"), val = int32(2)];
+            int32 x_461_split_axis_0 = const()[name = string("x_461_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_461_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_461_split_cast_fp16_1 = split(axis = x_461_split_axis_0, num_splits = x_461_split_num_splits_0, x = input_935_cast_fp16)[name = string("x_461_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_461_split_1_sigmoid_cast_fp16 = sigmoid(x = x_461_split_cast_fp16_1)[name = string("x_461_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_461_cast_fp16 = mul(x = x_461_split_cast_fp16_0, y = x_461_split_1_sigmoid_cast_fp16)[name = string("x_461_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_937_cast_fp16 = select(a = var_44_to_fp16, b = x_461_cast_fp16, cond = var_575)[name = string("input_937_cast_fp16")];
+            bool new_x_71_interleave_0 = const()[name = string("new_x_71_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_71_cast_fp16 = concat(axis = var_59, interleave = new_x_71_interleave_0, values = (cache_71_cast_fp16, input_937_cast_fp16))[name = string("new_x_71_cast_fp16")];
+            tensor<int32, [3]> var_4209_begin_0 = const()[name = string("op_4209_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_4209_end_0 = const()[name = string("op_4209_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_4209_end_mask_0 = const()[name = string("op_4209_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4209_cast_fp16 = slice_by_index(begin = var_4209_begin_0, end = var_4209_end_0, end_mask = var_4209_end_mask_0, x = new_x_71_cast_fp16)[name = string("op_4209_cast_fp16")];
+            string x_463_pad_type_0 = const()[name = string("x_463_pad_type_0"), val = string("valid")];
+            int32 x_463_groups_0 = const()[name = string("x_463_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_463_strides_0 = const()[name = string("x_463_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_463_pad_0 = const()[name = string("x_463_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_463_dilations_0 = const()[name = string("x_463_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353617536))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353626816))))[name = string("encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_463_cast_fp16 = conv(dilations = x_463_dilations_0, groups = x_463_groups_0, pad = x_463_pad_0, pad_type = x_463_pad_type_0, strides = x_463_strides_0, weight = encoder_layers_17_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_71_cast_fp16)[name = string("x_463_cast_fp16")];
+            tensor<int32, [3]> input_939_perm_0 = const()[name = string("input_939_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_465_axes_0 = const()[name = string("x_465_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353628928)))];
+            tensor<fp16, [1024]> encoder_layers_17_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_17_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353631040)))];
+            tensor<fp16, [1, 7, 1024]> input_939_cast_fp16 = transpose(perm = input_939_perm_0, x = x_463_cast_fp16)[name = string("transpose_203")];
+            tensor<fp16, [1, 7, 1024]> x_465_cast_fp16 = layer_norm(axes = x_465_axes_0, beta = encoder_layers_17_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_conv_batch_norm_weight_to_fp16, x = input_939_cast_fp16)[name = string("x_465_cast_fp16")];
+            tensor<int32, [3]> input_941_perm_0 = const()[name = string("input_941_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_941_cast_fp16 = transpose(perm = input_941_perm_0, x = x_465_cast_fp16)[name = string("transpose_202")];
+            tensor<fp16, [1, 1024, 7]> input_943_cast_fp16 = silu(x = input_941_cast_fp16)[name = string("input_943_cast_fp16")];
+            string x_467_pad_type_0 = const()[name = string("x_467_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_467_strides_0 = const()[name = string("x_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_467_pad_0 = const()[name = string("x_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_467_dilations_0 = const()[name = string("x_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_467_groups_0 = const()[name = string("x_467_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353633152))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354681792))))[name = string("encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_467_cast_fp16 = conv(dilations = x_467_dilations_0, groups = x_467_groups_0, pad = x_467_pad_0, pad_type = x_467_pad_type_0, strides = x_467_strides_0, weight = encoder_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_943_cast_fp16)[name = string("x_467_cast_fp16")];
+            tensor<int32, [3]> input_945_perm_0 = const()[name = string("input_945_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_945_cast_fp16 = transpose(perm = input_945_perm_0, x = x_467_cast_fp16)[name = string("transpose_201")];
+            tensor<fp16, [1, 7, 1024]> input_947_cast_fp16 = add(x = input_931_cast_fp16, y = input_945_cast_fp16)[name = string("input_947_cast_fp16")];
+            tensor<int32, [1]> input_949_axes_0 = const()[name = string("input_949_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354683904)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354686016)))];
+            tensor<fp16, [1, 7, 1024]> input_949_cast_fp16 = layer_norm(axes = input_949_axes_0, beta = encoder_layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_feed_forward2_weight_to_fp16, x = input_947_cast_fp16)[name = string("input_949_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354688128))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357833920))))[name = string("encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_17_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357834112)))];
+            tensor<fp16, [1, 7, 4096]> linear_161_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear1_weight_to_fp16_palettized, x = input_949_cast_fp16)[name = string("linear_161_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_953_cast_fp16 = silu(x = linear_161_cast_fp16)[name = string("input_953_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(357842368))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360988160))))[name = string("encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_17_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_17_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360988352)))];
+            tensor<fp16, [1, 7, 1024]> linear_162_cast_fp16 = linear(bias = encoder_layers_17_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_17_feed_forward2_linear2_weight_to_fp16_palettized, x = input_953_cast_fp16)[name = string("linear_162_cast_fp16")];
+            fp16 var_4252_to_fp16 = const()[name = string("op_4252_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4253_cast_fp16 = mul(x = linear_162_cast_fp16, y = var_4252_to_fp16)[name = string("op_4253_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_959_cast_fp16 = add(x = input_947_cast_fp16, y = var_4253_cast_fp16)[name = string("input_959_cast_fp16")];
+            tensor<int32, [1]> input_961_axes_0 = const()[name = string("input_961_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_17_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_17_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360990464)))];
+            tensor<fp16, [1024]> encoder_layers_17_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_17_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360992576)))];
+            tensor<fp16, [1, 7, 1024]> input_961_cast_fp16 = layer_norm(axes = input_961_axes_0, beta = encoder_layers_17_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_17_norm_out_weight_to_fp16, x = input_959_cast_fp16)[name = string("input_961_cast_fp16")];
+            tensor<int32, [4]> cache_73_begin_0 = const()[name = string("cache_73_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> cache_73_end_0 = const()[name = string("cache_73_end_0"), val = tensor<int32, [4]>([19, 1, 42, 1024])];
+            tensor<bool, [4]> cache_73_end_mask_0 = const()[name = string("cache_73_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_73_squeeze_mask_0 = const()[name = string("cache_73_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_73_cast_fp16 = slice_by_index(begin = cache_73_begin_0, end = cache_73_end_0, end_mask = cache_73_end_mask_0, squeeze_mask = cache_73_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_73_cast_fp16")];
+            tensor<int32, [4]> cache_75_begin_0 = const()[name = string("cache_75_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> cache_75_end_0 = const()[name = string("cache_75_end_0"), val = tensor<int32, [4]>([19, 1, 1024, 8])];
+            tensor<bool, [4]> cache_75_end_mask_0 = const()[name = string("cache_75_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_75_squeeze_mask_0 = const()[name = string("cache_75_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_75_cast_fp16 = slice_by_index(begin = cache_75_begin_0, end = cache_75_end_0, end_mask = cache_75_end_mask_0, squeeze_mask = cache_75_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_75_cast_fp16")];
+            tensor<int32, [1]> input_963_axes_0 = const()[name = string("input_963_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360994688)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360996800)))];
+            tensor<fp16, [1, 7, 1024]> input_963_cast_fp16 = layer_norm(axes = input_963_axes_0, beta = encoder_layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_feed_forward1_weight_to_fp16, x = input_961_cast_fp16)[name = string("input_963_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360998912))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364144704))))[name = string("encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized")];
+            tensor<fp16, [4096]> encoder_layers_18_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364144896)))];
+            tensor<fp16, [1, 7, 4096]> linear_163_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear1_weight_to_fp16_palettized, x = input_963_cast_fp16)[name = string("linear_163_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_967_cast_fp16 = silu(x = linear_163_cast_fp16)[name = string("input_967_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364153152))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367298944))))[name = string("encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367299136)))];
+            tensor<fp16, [1, 7, 1024]> linear_164_cast_fp16 = linear(bias = encoder_layers_18_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward1_linear2_weight_to_fp16_palettized, x = input_967_cast_fp16)[name = string("linear_164_cast_fp16")];
+            fp16 var_4289_to_fp16 = const()[name = string("op_4289_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4290_cast_fp16 = mul(x = linear_164_cast_fp16, y = var_4289_to_fp16)[name = string("op_4290_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_973_cast_fp16 = add(x = input_961_cast_fp16, y = var_4290_cast_fp16)[name = string("input_973_cast_fp16")];
+            tensor<int32, [1]> key_37_axes_0 = const()[name = string("key_37_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367301248)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367303360)))];
+            tensor<fp16, [1, 7, 1024]> key_37_cast_fp16 = layer_norm(axes = key_37_axes_0, beta = encoder_layers_18_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_self_att_weight_to_fp16, x = input_973_cast_fp16)[name = string("key_37_cast_fp16")];
+            bool input_975_interleave_0 = const()[name = string("input_975_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_975_cast_fp16 = concat(axis = var_68, interleave = input_975_interleave_0, values = (cache_73_cast_fp16, key_37_cast_fp16))[name = string("input_975_cast_fp16")];
+            tensor<int32, [3]> var_4312_begin_0 = const()[name = string("op_4312_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_4312_end_0 = const()[name = string("op_4312_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4312_end_mask_0 = const()[name = string("op_4312_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_4312_cast_fp16 = slice_by_index(begin = var_4312_begin_0, end = var_4312_end_0, end_mask = var_4312_end_mask_0, x = cache_73_cast_fp16)[name = string("op_4312_cast_fp16")];
+            bool var_4318_interleave_0 = const()[name = string("op_4318_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4318_cast_fp16 = concat(axis = var_68, interleave = var_4318_interleave_0, values = (var_4312_cast_fp16, key_37_cast_fp16))[name = string("op_4318_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367305472))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368091968))))[name = string("encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368092160)))];
+            tensor<fp16, [1, 7, 1024]> linear_165_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_q_weight_to_fp16_palettized, x = key_37_cast_fp16)[name = string("linear_165_cast_fp16")];
+            tensor<int32, [4]> var_4323 = const()[name = string("op_4323"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_109_cast_fp16 = reshape(shape = var_4323, x = linear_165_cast_fp16)[name = string("q_109_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368094272))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368880768))))[name = string("encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368880960)))];
+            tensor<fp16, [1, 49, 1024]> linear_166_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_k_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_166_cast_fp16")];
+            tensor<int32, [4]> var_4328 = const()[name = string("op_4328"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_73_cast_fp16 = reshape(shape = var_4328, x = linear_166_cast_fp16)[name = string("k_73_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(368883072))), lut = tensor<fp16, [1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369669568))))[name = string("encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369669760)))];
+            tensor<fp16, [1, 49, 1024]> linear_167_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_v_weight_to_fp16_palettized, x = input_975_cast_fp16)[name = string("linear_167_cast_fp16")];
+            tensor<int32, [4]> var_4333 = const()[name = string("op_4333"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_37_cast_fp16 = reshape(shape = var_4333, x = linear_167_cast_fp16)[name = string("v_37_cast_fp16")];
+            tensor<int32, [4]> value_45_perm_0 = const()[name = string("value_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_18_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369671872)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4346_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_u_to_fp16)[name = string("op_4346_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_18_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_18_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369673984)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4348_cast_fp16 = add(x = q_109_cast_fp16, y = encoder_layers_18_self_attn_pos_bias_v_to_fp16)[name = string("op_4348_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_37_perm_0 = const()[name = string("q_with_bias_v_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_475_transpose_x_0 = const()[name = string("x_475_transpose_x_0"), val = bool(false)];
+            bool x_475_transpose_y_0 = const()[name = string("x_475_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_4350_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369676096))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369775488))))[name = string("op_4350_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_37_cast_fp16 = transpose(perm = q_with_bias_v_37_perm_0, x = var_4348_cast_fp16)[name = string("transpose_200")];
+            tensor<fp16, [1, 8, 7, 97]> x_475_cast_fp16 = matmul(transpose_x = x_475_transpose_x_0, transpose_y = x_475_transpose_y_0, x = q_with_bias_v_37_cast_fp16, y = op_4350_to_fp16_quantized)[name = string("x_475_cast_fp16")];
+            tensor<int32, [8]> x_477_pad_0 = const()[name = string("x_477_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_477_mode_0 = const()[name = string("x_477_mode_0"), val = string("constant")];
+            fp16 const_313_to_fp16 = const()[name = string("const_313_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_477_cast_fp16 = pad(constant_val = const_313_to_fp16, mode = x_477_mode_0, pad = x_477_pad_0, x = x_475_cast_fp16)[name = string("x_477_cast_fp16")];
+            tensor<int32, [4]> var_4358 = const()[name = string("op_4358"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_479_cast_fp16 = reshape(shape = var_4358, x = x_477_cast_fp16)[name = string("x_479_cast_fp16")];
+            tensor<int32, [4]> var_4362_begin_0 = const()[name = string("op_4362_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4362_end_0 = const()[name = string("op_4362_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_4362_end_mask_0 = const()[name = string("op_4362_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_4362_cast_fp16 = slice_by_index(begin = var_4362_begin_0, end = var_4362_end_0, end_mask = var_4362_end_mask_0, x = x_479_cast_fp16)[name = string("op_4362_cast_fp16")];
+            tensor<int32, [4]> var_4363 = const()[name = string("op_4363"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_73_cast_fp16 = reshape(shape = var_4363, x = var_4362_cast_fp16)[name = string("matrix_bd_73_cast_fp16")];
+            bool matrix_ac_37_transpose_x_0 = const()[name = string("matrix_ac_37_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_37_transpose_y_0 = const()[name = string("matrix_ac_37_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_132_perm_0 = const()[name = string("transpose_132_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_133_perm_0 = const()[name = string("transpose_133_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_133 = transpose(perm = transpose_133_perm_0, x = k_73_cast_fp16)[name = string("transpose_198")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_132 = transpose(perm = transpose_132_perm_0, x = var_4346_cast_fp16)[name = string("transpose_199")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_37_cast_fp16 = matmul(transpose_x = matrix_ac_37_transpose_x_0, transpose_y = matrix_ac_37_transpose_y_0, x = transpose_132, y = transpose_133)[name = string("matrix_ac_37_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_75_begin_0 = const()[name = string("matrix_bd_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_75_end_0 = const()[name = string("matrix_bd_75_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_75_end_mask_0 = const()[name = string("matrix_bd_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_75_cast_fp16 = slice_by_index(begin = matrix_bd_75_begin_0, end = matrix_bd_75_end_0, end_mask = matrix_bd_75_end_mask_0, x = matrix_bd_73_cast_fp16)[name = string("matrix_bd_75_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_4372_cast_fp16 = add(x = matrix_ac_37_cast_fp16, y = matrix_bd_75_cast_fp16)[name = string("op_4372_cast_fp16")];
+            fp16 _inversed_scores_73_y_0_to_fp16 = const()[name = string("_inversed_scores_73_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_73_cast_fp16 = mul(x = var_4372_cast_fp16, y = _inversed_scores_73_y_0_to_fp16)[name = string("_inversed_scores_73_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_75_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_73_cast_fp16, cond = mask_11)[name = string("scores_75_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_4378_cast_fp16 = softmax(axis = var_59, x = scores_75_cast_fp16)[name = string("op_4378_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_977_cast_fp16 = select(a = var_44_to_fp16, b = var_4378_cast_fp16, cond = mask_11)[name = string("input_977_cast_fp16")];
+            bool x_481_transpose_x_0 = const()[name = string("x_481_transpose_x_0"), val = bool(false)];
+            bool x_481_transpose_y_0 = const()[name = string("x_481_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_45_cast_fp16 = transpose(perm = value_45_perm_0, x = v_37_cast_fp16)[name = string("transpose_197")];
+            tensor<fp16, [1, 8, 7, 128]> x_481_cast_fp16 = matmul(transpose_x = x_481_transpose_x_0, transpose_y = x_481_transpose_y_0, x = input_977_cast_fp16, y = value_45_cast_fp16)[name = string("x_481_cast_fp16")];
+            tensor<int32, [4]> var_4382_perm_0 = const()[name = string("op_4382_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4383 = const()[name = string("op_4383"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_4382_cast_fp16 = transpose(perm = var_4382_perm_0, x = x_481_cast_fp16)[name = string("transpose_196")];
+            tensor<fp16, [1, 7, 1024]> input_979_cast_fp16 = reshape(shape = var_4383, x = var_4382_cast_fp16)[name = string("input_979_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369775808))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370824448))))[name = string("encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_18_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_18_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370826560)))];
+            tensor<fp16, [1, 7, 1024]> linear_169_cast_fp16 = linear(bias = encoder_layers_18_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_18_self_attn_linear_out_weight_to_fp16_quantized, x = input_979_cast_fp16)[name = string("linear_169_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_983_cast_fp16 = add(x = input_973_cast_fp16, y = linear_169_cast_fp16)[name = string("input_983_cast_fp16")];
+            tensor<int32, [1]> x_485_axes_0 = const()[name = string("x_485_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370828672)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370830784)))];
+            tensor<fp16, [1, 7, 1024]> x_485_cast_fp16 = layer_norm(axes = x_485_axes_0, beta = encoder_layers_18_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_conv_weight_to_fp16, x = input_983_cast_fp16)[name = string("x_485_cast_fp16")];
+            tensor<int32, [3]> input_985_perm_0 = const()[name = string("input_985_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_987_pad_type_0 = const()[name = string("input_987_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_987_strides_0 = const()[name = string("input_987_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_987_pad_0 = const()[name = string("input_987_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_987_dilations_0 = const()[name = string("input_987_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_987_groups_0 = const()[name = string("input_987_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(370832896))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372930112))))[name = string("encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_985_cast_fp16 = transpose(perm = input_985_perm_0, x = x_485_cast_fp16)[name = string("transpose_195")];
+            tensor<fp16, [1, 2048, 7]> input_987_cast_fp16 = conv(dilations = input_987_dilations_0, groups = input_987_groups_0, pad = input_987_pad_0, pad_type = input_987_pad_type_0, strides = input_987_strides_0, weight = encoder_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_985_cast_fp16)[name = string("input_987_cast_fp16")];
+            int32 x_487_split_num_splits_0 = const()[name = string("x_487_split_num_splits_0"), val = int32(2)];
+            int32 x_487_split_axis_0 = const()[name = string("x_487_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_487_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_487_split_cast_fp16_1 = split(axis = x_487_split_axis_0, num_splits = x_487_split_num_splits_0, x = input_987_cast_fp16)[name = string("x_487_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_487_split_1_sigmoid_cast_fp16 = sigmoid(x = x_487_split_cast_fp16_1)[name = string("x_487_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_487_cast_fp16 = mul(x = x_487_split_cast_fp16_0, y = x_487_split_1_sigmoid_cast_fp16)[name = string("x_487_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_989_cast_fp16 = select(a = var_44_to_fp16, b = x_487_cast_fp16, cond = var_575)[name = string("input_989_cast_fp16")];
+            bool new_x_75_interleave_0 = const()[name = string("new_x_75_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_75_cast_fp16 = concat(axis = var_59, interleave = new_x_75_interleave_0, values = (cache_75_cast_fp16, input_989_cast_fp16))[name = string("new_x_75_cast_fp16")];
+            tensor<int32, [3]> var_4422_begin_0 = const()[name = string("op_4422_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_4422_end_0 = const()[name = string("op_4422_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_4422_end_mask_0 = const()[name = string("op_4422_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4422_cast_fp16 = slice_by_index(begin = var_4422_begin_0, end = var_4422_end_0, end_mask = var_4422_end_mask_0, x = new_x_75_cast_fp16)[name = string("op_4422_cast_fp16")];
+            string x_489_pad_type_0 = const()[name = string("x_489_pad_type_0"), val = string("valid")];
+            int32 x_489_groups_0 = const()[name = string("x_489_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_489_strides_0 = const()[name = string("x_489_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_489_pad_0 = const()[name = string("x_489_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_489_dilations_0 = const()[name = string("x_489_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372934272))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372943552))))[name = string("encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_489_cast_fp16 = conv(dilations = x_489_dilations_0, groups = x_489_groups_0, pad = x_489_pad_0, pad_type = x_489_pad_type_0, strides = x_489_strides_0, weight = encoder_layers_18_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_75_cast_fp16)[name = string("x_489_cast_fp16")];
+            tensor<int32, [3]> input_991_perm_0 = const()[name = string("input_991_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_491_axes_0 = const()[name = string("x_491_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372945664)))];
+            tensor<fp16, [1024]> encoder_layers_18_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_18_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372947776)))];
+            tensor<fp16, [1, 7, 1024]> input_991_cast_fp16 = transpose(perm = input_991_perm_0, x = x_489_cast_fp16)[name = string("transpose_194")];
+            tensor<fp16, [1, 7, 1024]> x_491_cast_fp16 = layer_norm(axes = x_491_axes_0, beta = encoder_layers_18_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_conv_batch_norm_weight_to_fp16, x = input_991_cast_fp16)[name = string("x_491_cast_fp16")];
+            tensor<int32, [3]> input_993_perm_0 = const()[name = string("input_993_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_993_cast_fp16 = transpose(perm = input_993_perm_0, x = x_491_cast_fp16)[name = string("transpose_193")];
+            tensor<fp16, [1, 1024, 7]> input_995_cast_fp16 = silu(x = input_993_cast_fp16)[name = string("input_995_cast_fp16")];
+            string x_493_pad_type_0 = const()[name = string("x_493_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_493_strides_0 = const()[name = string("x_493_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_493_pad_0 = const()[name = string("x_493_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_493_dilations_0 = const()[name = string("x_493_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_493_groups_0 = const()[name = string("x_493_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372949888))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(373998528))))[name = string("encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_493_cast_fp16 = conv(dilations = x_493_dilations_0, groups = x_493_groups_0, pad = x_493_pad_0, pad_type = x_493_pad_type_0, strides = x_493_strides_0, weight = encoder_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_995_cast_fp16)[name = string("x_493_cast_fp16")];
+            tensor<int32, [3]> input_997_perm_0 = const()[name = string("input_997_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_997_cast_fp16 = transpose(perm = input_997_perm_0, x = x_493_cast_fp16)[name = string("transpose_192")];
+            tensor<fp16, [1, 7, 1024]> input_999_cast_fp16 = add(x = input_983_cast_fp16, y = input_997_cast_fp16)[name = string("input_999_cast_fp16")];
+            tensor<int32, [1]> input_1001_axes_0 = const()[name = string("input_1001_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374000640)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374002752)))];
+            tensor<fp16, [1, 7, 1024]> input_1001_cast_fp16 = layer_norm(axes = input_1001_axes_0, beta = encoder_layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_feed_forward2_weight_to_fp16, x = input_999_cast_fp16)[name = string("input_1001_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374004864))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378199232))))[name = string("encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_18_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378207488)))];
+            tensor<fp16, [1, 7, 4096]> linear_170_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1001_cast_fp16)[name = string("linear_170_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1005_cast_fp16 = silu(x = linear_170_cast_fp16)[name = string("input_1005_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378215744))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382410112))))[name = string("encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_18_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_18_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382412224)))];
+            tensor<fp16, [1, 7, 1024]> linear_171_cast_fp16 = linear(bias = encoder_layers_18_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_18_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1005_cast_fp16)[name = string("linear_171_cast_fp16")];
+            fp16 var_4465_to_fp16 = const()[name = string("op_4465_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4466_cast_fp16 = mul(x = linear_171_cast_fp16, y = var_4465_to_fp16)[name = string("op_4466_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1011_cast_fp16 = add(x = input_999_cast_fp16, y = var_4466_cast_fp16)[name = string("input_1011_cast_fp16")];
+            tensor<int32, [1]> input_1013_axes_0 = const()[name = string("input_1013_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_18_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_18_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382414336)))];
+            tensor<fp16, [1024]> encoder_layers_18_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_18_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382416448)))];
+            tensor<fp16, [1, 7, 1024]> input_1013_cast_fp16 = layer_norm(axes = input_1013_axes_0, beta = encoder_layers_18_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_18_norm_out_weight_to_fp16, x = input_1011_cast_fp16)[name = string("input_1013_cast_fp16")];
+            tensor<int32, [4]> cache_77_begin_0 = const()[name = string("cache_77_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
+            tensor<int32, [4]> cache_77_end_0 = const()[name = string("cache_77_end_0"), val = tensor<int32, [4]>([20, 1, 42, 1024])];
+            tensor<bool, [4]> cache_77_end_mask_0 = const()[name = string("cache_77_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_77_squeeze_mask_0 = const()[name = string("cache_77_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_77_cast_fp16 = slice_by_index(begin = cache_77_begin_0, end = cache_77_end_0, end_mask = cache_77_end_mask_0, squeeze_mask = cache_77_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_77_cast_fp16")];
+            tensor<int32, [4]> cache_79_begin_0 = const()[name = string("cache_79_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
+            tensor<int32, [4]> cache_79_end_0 = const()[name = string("cache_79_end_0"), val = tensor<int32, [4]>([20, 1, 1024, 8])];
+            tensor<bool, [4]> cache_79_end_mask_0 = const()[name = string("cache_79_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_79_squeeze_mask_0 = const()[name = string("cache_79_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_79_cast_fp16 = slice_by_index(begin = cache_79_begin_0, end = cache_79_end_0, end_mask = cache_79_end_mask_0, squeeze_mask = cache_79_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_79_cast_fp16")];
+            tensor<int32, [1]> input_1015_axes_0 = const()[name = string("input_1015_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382418560)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382420672)))];
+            tensor<fp16, [1, 7, 1024]> input_1015_cast_fp16 = layer_norm(axes = input_1015_axes_0, beta = encoder_layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_feed_forward1_weight_to_fp16, x = input_1013_cast_fp16)[name = string("input_1015_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382422784))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386617152))))[name = string("encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_19_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386625408)))];
+            tensor<fp16, [1, 7, 4096]> linear_172_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1015_cast_fp16)[name = string("linear_172_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1019_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("input_1019_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386633664))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390828032))))[name = string("encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390830144)))];
+            tensor<fp16, [1, 7, 1024]> linear_173_cast_fp16 = linear(bias = encoder_layers_19_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1019_cast_fp16)[name = string("linear_173_cast_fp16")];
+            fp16 var_4502_to_fp16 = const()[name = string("op_4502_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4503_cast_fp16 = mul(x = linear_173_cast_fp16, y = var_4502_to_fp16)[name = string("op_4503_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1025_cast_fp16 = add(x = input_1013_cast_fp16, y = var_4503_cast_fp16)[name = string("input_1025_cast_fp16")];
+            tensor<int32, [1]> key_39_axes_0 = const()[name = string("key_39_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390832256)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390834368)))];
+            tensor<fp16, [1, 7, 1024]> key_39_cast_fp16 = layer_norm(axes = key_39_axes_0, beta = encoder_layers_19_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_self_att_weight_to_fp16, x = input_1025_cast_fp16)[name = string("key_39_cast_fp16")];
+            bool input_1027_interleave_0 = const()[name = string("input_1027_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_1027_cast_fp16 = concat(axis = var_68, interleave = input_1027_interleave_0, values = (cache_77_cast_fp16, key_39_cast_fp16))[name = string("input_1027_cast_fp16")];
+            tensor<int32, [3]> var_4525_begin_0 = const()[name = string("op_4525_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_4525_end_0 = const()[name = string("op_4525_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4525_end_mask_0 = const()[name = string("op_4525_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_4525_cast_fp16 = slice_by_index(begin = var_4525_begin_0, end = var_4525_end_0, end_mask = var_4525_end_mask_0, x = cache_77_cast_fp16)[name = string("op_4525_cast_fp16")];
+            bool var_4531_interleave_0 = const()[name = string("op_4531_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4531_cast_fp16 = concat(axis = var_68, interleave = var_4531_interleave_0, values = (var_4525_cast_fp16, key_39_cast_fp16))[name = string("op_4531_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390836480))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391885120))))[name = string("encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391887232)))];
+            tensor<fp16, [1, 7, 1024]> linear_174_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_q_weight_to_fp16_quantized, x = key_39_cast_fp16)[name = string("linear_174_cast_fp16")];
+            tensor<int32, [4]> var_4536 = const()[name = string("op_4536"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_115_cast_fp16 = reshape(shape = var_4536, x = linear_174_cast_fp16)[name = string("q_115_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391889344))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392937984))))[name = string("encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392940096)))];
+            tensor<fp16, [1, 49, 1024]> linear_175_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_k_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_175_cast_fp16")];
+            tensor<int32, [4]> var_4541 = const()[name = string("op_4541"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_77_cast_fp16 = reshape(shape = var_4541, x = linear_175_cast_fp16)[name = string("k_77_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392942208))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393990848))))[name = string("encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393992960)))];
+            tensor<fp16, [1, 49, 1024]> linear_176_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_v_weight_to_fp16_quantized, x = input_1027_cast_fp16)[name = string("linear_176_cast_fp16")];
+            tensor<int32, [4]> var_4546 = const()[name = string("op_4546"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_39_cast_fp16 = reshape(shape = var_4546, x = linear_176_cast_fp16)[name = string("v_39_cast_fp16")];
+            tensor<int32, [4]> value_47_perm_0 = const()[name = string("value_47_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_19_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393995072)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4559_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_u_to_fp16)[name = string("op_4559_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_19_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_19_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393997184)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4561_cast_fp16 = add(x = q_115_cast_fp16, y = encoder_layers_19_self_attn_pos_bias_v_to_fp16)[name = string("op_4561_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_39_perm_0 = const()[name = string("q_with_bias_v_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_501_transpose_x_0 = const()[name = string("x_501_transpose_x_0"), val = bool(false)];
+            bool x_501_transpose_y_0 = const()[name = string("x_501_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_4563_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393999296))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394098688))))[name = string("op_4563_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_39_cast_fp16 = transpose(perm = q_with_bias_v_39_perm_0, x = var_4561_cast_fp16)[name = string("transpose_191")];
+            tensor<fp16, [1, 8, 7, 97]> x_501_cast_fp16 = matmul(transpose_x = x_501_transpose_x_0, transpose_y = x_501_transpose_y_0, x = q_with_bias_v_39_cast_fp16, y = op_4563_to_fp16_quantized)[name = string("x_501_cast_fp16")];
+            tensor<int32, [8]> x_503_pad_0 = const()[name = string("x_503_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_503_mode_0 = const()[name = string("x_503_mode_0"), val = string("constant")];
+            fp16 const_326_to_fp16 = const()[name = string("const_326_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_503_cast_fp16 = pad(constant_val = const_326_to_fp16, mode = x_503_mode_0, pad = x_503_pad_0, x = x_501_cast_fp16)[name = string("x_503_cast_fp16")];
+            tensor<int32, [4]> var_4571 = const()[name = string("op_4571"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_505_cast_fp16 = reshape(shape = var_4571, x = x_503_cast_fp16)[name = string("x_505_cast_fp16")];
+            tensor<int32, [4]> var_4575_begin_0 = const()[name = string("op_4575_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4575_end_0 = const()[name = string("op_4575_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_4575_end_mask_0 = const()[name = string("op_4575_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_4575_cast_fp16 = slice_by_index(begin = var_4575_begin_0, end = var_4575_end_0, end_mask = var_4575_end_mask_0, x = x_505_cast_fp16)[name = string("op_4575_cast_fp16")];
+            tensor<int32, [4]> var_4576 = const()[name = string("op_4576"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_77_cast_fp16 = reshape(shape = var_4576, x = var_4575_cast_fp16)[name = string("matrix_bd_77_cast_fp16")];
+            bool matrix_ac_39_transpose_x_0 = const()[name = string("matrix_ac_39_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_39_transpose_y_0 = const()[name = string("matrix_ac_39_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_134_perm_0 = const()[name = string("transpose_134_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_135_perm_0 = const()[name = string("transpose_135_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_135 = transpose(perm = transpose_135_perm_0, x = k_77_cast_fp16)[name = string("transpose_189")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_134 = transpose(perm = transpose_134_perm_0, x = var_4559_cast_fp16)[name = string("transpose_190")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_39_cast_fp16 = matmul(transpose_x = matrix_ac_39_transpose_x_0, transpose_y = matrix_ac_39_transpose_y_0, x = transpose_134, y = transpose_135)[name = string("matrix_ac_39_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_79_begin_0 = const()[name = string("matrix_bd_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_79_end_0 = const()[name = string("matrix_bd_79_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_79_end_mask_0 = const()[name = string("matrix_bd_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_79_cast_fp16 = slice_by_index(begin = matrix_bd_79_begin_0, end = matrix_bd_79_end_0, end_mask = matrix_bd_79_end_mask_0, x = matrix_bd_77_cast_fp16)[name = string("matrix_bd_79_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_4585_cast_fp16 = add(x = matrix_ac_39_cast_fp16, y = matrix_bd_79_cast_fp16)[name = string("op_4585_cast_fp16")];
+            fp16 _inversed_scores_77_y_0_to_fp16 = const()[name = string("_inversed_scores_77_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_77_cast_fp16 = mul(x = var_4585_cast_fp16, y = _inversed_scores_77_y_0_to_fp16)[name = string("_inversed_scores_77_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_79_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_77_cast_fp16, cond = mask_11)[name = string("scores_79_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_4591_cast_fp16 = softmax(axis = var_59, x = scores_79_cast_fp16)[name = string("op_4591_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_1029_cast_fp16 = select(a = var_44_to_fp16, b = var_4591_cast_fp16, cond = mask_11)[name = string("input_1029_cast_fp16")];
+            bool x_507_transpose_x_0 = const()[name = string("x_507_transpose_x_0"), val = bool(false)];
+            bool x_507_transpose_y_0 = const()[name = string("x_507_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_47_cast_fp16 = transpose(perm = value_47_perm_0, x = v_39_cast_fp16)[name = string("transpose_188")];
+            tensor<fp16, [1, 8, 7, 128]> x_507_cast_fp16 = matmul(transpose_x = x_507_transpose_x_0, transpose_y = x_507_transpose_y_0, x = input_1029_cast_fp16, y = value_47_cast_fp16)[name = string("x_507_cast_fp16")];
+            tensor<int32, [4]> var_4595_perm_0 = const()[name = string("op_4595_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4596 = const()[name = string("op_4596"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_4595_cast_fp16 = transpose(perm = var_4595_perm_0, x = x_507_cast_fp16)[name = string("transpose_187")];
+            tensor<fp16, [1, 7, 1024]> input_1031_cast_fp16 = reshape(shape = var_4596, x = var_4595_cast_fp16)[name = string("input_1031_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394099008))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395147648))))[name = string("encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_19_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395149760)))];
+            tensor<fp16, [1, 7, 1024]> linear_178_cast_fp16 = linear(bias = encoder_layers_19_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_19_self_attn_linear_out_weight_to_fp16_quantized, x = input_1031_cast_fp16)[name = string("linear_178_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1035_cast_fp16 = add(x = input_1025_cast_fp16, y = linear_178_cast_fp16)[name = string("input_1035_cast_fp16")];
+            tensor<int32, [1]> x_511_axes_0 = const()[name = string("x_511_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395151872)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395153984)))];
+            tensor<fp16, [1, 7, 1024]> x_511_cast_fp16 = layer_norm(axes = x_511_axes_0, beta = encoder_layers_19_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_conv_weight_to_fp16, x = input_1035_cast_fp16)[name = string("x_511_cast_fp16")];
+            tensor<int32, [3]> input_1037_perm_0 = const()[name = string("input_1037_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1039_pad_type_0 = const()[name = string("input_1039_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1039_strides_0 = const()[name = string("input_1039_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1039_pad_0 = const()[name = string("input_1039_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1039_dilations_0 = const()[name = string("input_1039_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1039_groups_0 = const()[name = string("input_1039_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(395156096))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397253312))))[name = string("encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_1037_cast_fp16 = transpose(perm = input_1037_perm_0, x = x_511_cast_fp16)[name = string("transpose_186")];
+            tensor<fp16, [1, 2048, 7]> input_1039_cast_fp16 = conv(dilations = input_1039_dilations_0, groups = input_1039_groups_0, pad = input_1039_pad_0, pad_type = input_1039_pad_type_0, strides = input_1039_strides_0, weight = encoder_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1037_cast_fp16)[name = string("input_1039_cast_fp16")];
+            int32 x_513_split_num_splits_0 = const()[name = string("x_513_split_num_splits_0"), val = int32(2)];
+            int32 x_513_split_axis_0 = const()[name = string("x_513_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_513_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_513_split_cast_fp16_1 = split(axis = x_513_split_axis_0, num_splits = x_513_split_num_splits_0, x = input_1039_cast_fp16)[name = string("x_513_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_513_split_1_sigmoid_cast_fp16 = sigmoid(x = x_513_split_cast_fp16_1)[name = string("x_513_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_513_cast_fp16 = mul(x = x_513_split_cast_fp16_0, y = x_513_split_1_sigmoid_cast_fp16)[name = string("x_513_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_1041_cast_fp16 = select(a = var_44_to_fp16, b = x_513_cast_fp16, cond = var_575)[name = string("input_1041_cast_fp16")];
+            bool new_x_79_interleave_0 = const()[name = string("new_x_79_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_79_cast_fp16 = concat(axis = var_59, interleave = new_x_79_interleave_0, values = (cache_79_cast_fp16, input_1041_cast_fp16))[name = string("new_x_79_cast_fp16")];
+            tensor<int32, [3]> var_4635_begin_0 = const()[name = string("op_4635_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_4635_end_0 = const()[name = string("op_4635_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_4635_end_mask_0 = const()[name = string("op_4635_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4635_cast_fp16 = slice_by_index(begin = var_4635_begin_0, end = var_4635_end_0, end_mask = var_4635_end_mask_0, x = new_x_79_cast_fp16)[name = string("op_4635_cast_fp16")];
+            string x_515_pad_type_0 = const()[name = string("x_515_pad_type_0"), val = string("valid")];
+            int32 x_515_groups_0 = const()[name = string("x_515_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_515_strides_0 = const()[name = string("x_515_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_515_pad_0 = const()[name = string("x_515_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_515_dilations_0 = const()[name = string("x_515_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397257472))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397266752))))[name = string("encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_515_cast_fp16 = conv(dilations = x_515_dilations_0, groups = x_515_groups_0, pad = x_515_pad_0, pad_type = x_515_pad_type_0, strides = x_515_strides_0, weight = encoder_layers_19_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_79_cast_fp16)[name = string("x_515_cast_fp16")];
+            tensor<int32, [3]> input_1043_perm_0 = const()[name = string("input_1043_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_517_axes_0 = const()[name = string("x_517_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397268864)))];
+            tensor<fp16, [1024]> encoder_layers_19_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_19_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397270976)))];
+            tensor<fp16, [1, 7, 1024]> input_1043_cast_fp16 = transpose(perm = input_1043_perm_0, x = x_515_cast_fp16)[name = string("transpose_185")];
+            tensor<fp16, [1, 7, 1024]> x_517_cast_fp16 = layer_norm(axes = x_517_axes_0, beta = encoder_layers_19_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_conv_batch_norm_weight_to_fp16, x = input_1043_cast_fp16)[name = string("x_517_cast_fp16")];
+            tensor<int32, [3]> input_1045_perm_0 = const()[name = string("input_1045_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_1045_cast_fp16 = transpose(perm = input_1045_perm_0, x = x_517_cast_fp16)[name = string("transpose_184")];
+            tensor<fp16, [1, 1024, 7]> input_1047_cast_fp16 = silu(x = input_1045_cast_fp16)[name = string("input_1047_cast_fp16")];
+            string x_519_pad_type_0 = const()[name = string("x_519_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_519_strides_0 = const()[name = string("x_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_519_pad_0 = const()[name = string("x_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_519_dilations_0 = const()[name = string("x_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_519_groups_0 = const()[name = string("x_519_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(397273088))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398321728))))[name = string("encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_519_cast_fp16 = conv(dilations = x_519_dilations_0, groups = x_519_groups_0, pad = x_519_pad_0, pad_type = x_519_pad_type_0, strides = x_519_strides_0, weight = encoder_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1047_cast_fp16)[name = string("x_519_cast_fp16")];
+            tensor<int32, [3]> input_1049_perm_0 = const()[name = string("input_1049_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_1049_cast_fp16 = transpose(perm = input_1049_perm_0, x = x_519_cast_fp16)[name = string("transpose_183")];
+            tensor<fp16, [1, 7, 1024]> input_1051_cast_fp16 = add(x = input_1035_cast_fp16, y = input_1049_cast_fp16)[name = string("input_1051_cast_fp16")];
+            tensor<int32, [1]> input_1053_axes_0 = const()[name = string("input_1053_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398323840)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398325952)))];
+            tensor<fp16, [1, 7, 1024]> input_1053_cast_fp16 = layer_norm(axes = input_1053_axes_0, beta = encoder_layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_feed_forward2_weight_to_fp16, x = input_1051_cast_fp16)[name = string("input_1053_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398328064))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402522432))))[name = string("encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_19_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402530688)))];
+            tensor<fp16, [1, 7, 4096]> linear_179_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1053_cast_fp16)[name = string("linear_179_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1057_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("input_1057_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402538944))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406733312))))[name = string("encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_19_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_19_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406735424)))];
+            tensor<fp16, [1, 7, 1024]> linear_180_cast_fp16 = linear(bias = encoder_layers_19_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_19_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1057_cast_fp16)[name = string("linear_180_cast_fp16")];
+            fp16 var_4678_to_fp16 = const()[name = string("op_4678_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4679_cast_fp16 = mul(x = linear_180_cast_fp16, y = var_4678_to_fp16)[name = string("op_4679_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1063_cast_fp16 = add(x = input_1051_cast_fp16, y = var_4679_cast_fp16)[name = string("input_1063_cast_fp16")];
+            tensor<int32, [1]> input_1065_axes_0 = const()[name = string("input_1065_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_19_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_19_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406737536)))];
+            tensor<fp16, [1024]> encoder_layers_19_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_19_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406739648)))];
+            tensor<fp16, [1, 7, 1024]> input_1065_cast_fp16 = layer_norm(axes = input_1065_axes_0, beta = encoder_layers_19_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_19_norm_out_weight_to_fp16, x = input_1063_cast_fp16)[name = string("input_1065_cast_fp16")];
+            tensor<int32, [4]> cache_81_begin_0 = const()[name = string("cache_81_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
+            tensor<int32, [4]> cache_81_end_0 = const()[name = string("cache_81_end_0"), val = tensor<int32, [4]>([21, 1, 42, 1024])];
+            tensor<bool, [4]> cache_81_end_mask_0 = const()[name = string("cache_81_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_81_squeeze_mask_0 = const()[name = string("cache_81_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_81_cast_fp16 = slice_by_index(begin = cache_81_begin_0, end = cache_81_end_0, end_mask = cache_81_end_mask_0, squeeze_mask = cache_81_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_81_cast_fp16")];
+            tensor<int32, [4]> cache_83_begin_0 = const()[name = string("cache_83_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
+            tensor<int32, [4]> cache_83_end_0 = const()[name = string("cache_83_end_0"), val = tensor<int32, [4]>([21, 1, 1024, 8])];
+            tensor<bool, [4]> cache_83_end_mask_0 = const()[name = string("cache_83_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_83_squeeze_mask_0 = const()[name = string("cache_83_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_83_cast_fp16 = slice_by_index(begin = cache_83_begin_0, end = cache_83_end_0, end_mask = cache_83_end_mask_0, squeeze_mask = cache_83_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_83_cast_fp16")];
+            tensor<int32, [1]> input_1067_axes_0 = const()[name = string("input_1067_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406741760)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406743872)))];
+            tensor<fp16, [1, 7, 1024]> input_1067_cast_fp16 = layer_norm(axes = input_1067_axes_0, beta = encoder_layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_feed_forward1_weight_to_fp16, x = input_1065_cast_fp16)[name = string("input_1067_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(406745984))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410940352))))[name = string("encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_20_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410948608)))];
+            tensor<fp16, [1, 7, 4096]> linear_181_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1067_cast_fp16)[name = string("linear_181_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1071_cast_fp16 = silu(x = linear_181_cast_fp16)[name = string("input_1071_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410956864))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415151232))))[name = string("encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415153344)))];
+            tensor<fp16, [1, 7, 1024]> linear_182_cast_fp16 = linear(bias = encoder_layers_20_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1071_cast_fp16)[name = string("linear_182_cast_fp16")];
+            fp16 var_4715_to_fp16 = const()[name = string("op_4715_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4716_cast_fp16 = mul(x = linear_182_cast_fp16, y = var_4715_to_fp16)[name = string("op_4716_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1077_cast_fp16 = add(x = input_1065_cast_fp16, y = var_4716_cast_fp16)[name = string("input_1077_cast_fp16")];
+            tensor<int32, [1]> key_41_axes_0 = const()[name = string("key_41_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415155456)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415157568)))];
+            tensor<fp16, [1, 7, 1024]> key_41_cast_fp16 = layer_norm(axes = key_41_axes_0, beta = encoder_layers_20_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_self_att_weight_to_fp16, x = input_1077_cast_fp16)[name = string("key_41_cast_fp16")];
+            bool input_1079_interleave_0 = const()[name = string("input_1079_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_1079_cast_fp16 = concat(axis = var_68, interleave = input_1079_interleave_0, values = (cache_81_cast_fp16, key_41_cast_fp16))[name = string("input_1079_cast_fp16")];
+            tensor<int32, [3]> var_4738_begin_0 = const()[name = string("op_4738_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_4738_end_0 = const()[name = string("op_4738_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4738_end_mask_0 = const()[name = string("op_4738_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_4738_cast_fp16 = slice_by_index(begin = var_4738_begin_0, end = var_4738_end_0, end_mask = var_4738_end_mask_0, x = cache_81_cast_fp16)[name = string("op_4738_cast_fp16")];
+            bool var_4744_interleave_0 = const()[name = string("op_4744_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4744_cast_fp16 = concat(axis = var_68, interleave = var_4744_interleave_0, values = (var_4738_cast_fp16, key_41_cast_fp16))[name = string("op_4744_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415159680))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416208320))))[name = string("encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416210432)))];
+            tensor<fp16, [1, 7, 1024]> linear_183_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_q_weight_to_fp16_quantized, x = key_41_cast_fp16)[name = string("linear_183_cast_fp16")];
+            tensor<int32, [4]> var_4749 = const()[name = string("op_4749"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_121_cast_fp16 = reshape(shape = var_4749, x = linear_183_cast_fp16)[name = string("q_121_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(416212544))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417261184))))[name = string("encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417263296)))];
+            tensor<fp16, [1, 49, 1024]> linear_184_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_k_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_184_cast_fp16")];
+            tensor<int32, [4]> var_4754 = const()[name = string("op_4754"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_81_cast_fp16 = reshape(shape = var_4754, x = linear_184_cast_fp16)[name = string("k_81_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(417265408))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418314048))))[name = string("encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418316160)))];
+            tensor<fp16, [1, 49, 1024]> linear_185_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_v_weight_to_fp16_quantized, x = input_1079_cast_fp16)[name = string("linear_185_cast_fp16")];
+            tensor<int32, [4]> var_4759 = const()[name = string("op_4759"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_41_cast_fp16 = reshape(shape = var_4759, x = linear_185_cast_fp16)[name = string("v_41_cast_fp16")];
+            tensor<int32, [4]> value_49_perm_0 = const()[name = string("value_49_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_20_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418318272)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4772_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_u_to_fp16)[name = string("op_4772_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_20_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_20_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418320384)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4774_cast_fp16 = add(x = q_121_cast_fp16, y = encoder_layers_20_self_attn_pos_bias_v_to_fp16)[name = string("op_4774_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_41_perm_0 = const()[name = string("q_with_bias_v_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_527_transpose_x_0 = const()[name = string("x_527_transpose_x_0"), val = bool(false)];
+            bool x_527_transpose_y_0 = const()[name = string("x_527_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_4776_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418322496))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418421888))))[name = string("op_4776_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_41_cast_fp16 = transpose(perm = q_with_bias_v_41_perm_0, x = var_4774_cast_fp16)[name = string("transpose_182")];
+            tensor<fp16, [1, 8, 7, 97]> x_527_cast_fp16 = matmul(transpose_x = x_527_transpose_x_0, transpose_y = x_527_transpose_y_0, x = q_with_bias_v_41_cast_fp16, y = op_4776_to_fp16_quantized)[name = string("x_527_cast_fp16")];
+            tensor<int32, [8]> x_529_pad_0 = const()[name = string("x_529_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_529_mode_0 = const()[name = string("x_529_mode_0"), val = string("constant")];
+            fp16 const_339_to_fp16 = const()[name = string("const_339_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_529_cast_fp16 = pad(constant_val = const_339_to_fp16, mode = x_529_mode_0, pad = x_529_pad_0, x = x_527_cast_fp16)[name = string("x_529_cast_fp16")];
+            tensor<int32, [4]> var_4784 = const()[name = string("op_4784"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_531_cast_fp16 = reshape(shape = var_4784, x = x_529_cast_fp16)[name = string("x_531_cast_fp16")];
+            tensor<int32, [4]> var_4788_begin_0 = const()[name = string("op_4788_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4788_end_0 = const()[name = string("op_4788_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_4788_end_mask_0 = const()[name = string("op_4788_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_4788_cast_fp16 = slice_by_index(begin = var_4788_begin_0, end = var_4788_end_0, end_mask = var_4788_end_mask_0, x = x_531_cast_fp16)[name = string("op_4788_cast_fp16")];
+            tensor<int32, [4]> var_4789 = const()[name = string("op_4789"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_81_cast_fp16 = reshape(shape = var_4789, x = var_4788_cast_fp16)[name = string("matrix_bd_81_cast_fp16")];
+            bool matrix_ac_41_transpose_x_0 = const()[name = string("matrix_ac_41_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_41_transpose_y_0 = const()[name = string("matrix_ac_41_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_136_perm_0 = const()[name = string("transpose_136_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_137_perm_0 = const()[name = string("transpose_137_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_137 = transpose(perm = transpose_137_perm_0, x = k_81_cast_fp16)[name = string("transpose_180")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_136 = transpose(perm = transpose_136_perm_0, x = var_4772_cast_fp16)[name = string("transpose_181")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_41_cast_fp16 = matmul(transpose_x = matrix_ac_41_transpose_x_0, transpose_y = matrix_ac_41_transpose_y_0, x = transpose_136, y = transpose_137)[name = string("matrix_ac_41_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_83_begin_0 = const()[name = string("matrix_bd_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_83_end_0 = const()[name = string("matrix_bd_83_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_83_end_mask_0 = const()[name = string("matrix_bd_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_83_cast_fp16 = slice_by_index(begin = matrix_bd_83_begin_0, end = matrix_bd_83_end_0, end_mask = matrix_bd_83_end_mask_0, x = matrix_bd_81_cast_fp16)[name = string("matrix_bd_83_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_4798_cast_fp16 = add(x = matrix_ac_41_cast_fp16, y = matrix_bd_83_cast_fp16)[name = string("op_4798_cast_fp16")];
+            fp16 _inversed_scores_81_y_0_to_fp16 = const()[name = string("_inversed_scores_81_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_81_cast_fp16 = mul(x = var_4798_cast_fp16, y = _inversed_scores_81_y_0_to_fp16)[name = string("_inversed_scores_81_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_83_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_81_cast_fp16, cond = mask_11)[name = string("scores_83_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_4804_cast_fp16 = softmax(axis = var_59, x = scores_83_cast_fp16)[name = string("op_4804_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_1081_cast_fp16 = select(a = var_44_to_fp16, b = var_4804_cast_fp16, cond = mask_11)[name = string("input_1081_cast_fp16")];
+            bool x_533_transpose_x_0 = const()[name = string("x_533_transpose_x_0"), val = bool(false)];
+            bool x_533_transpose_y_0 = const()[name = string("x_533_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_49_cast_fp16 = transpose(perm = value_49_perm_0, x = v_41_cast_fp16)[name = string("transpose_179")];
+            tensor<fp16, [1, 8, 7, 128]> x_533_cast_fp16 = matmul(transpose_x = x_533_transpose_x_0, transpose_y = x_533_transpose_y_0, x = input_1081_cast_fp16, y = value_49_cast_fp16)[name = string("x_533_cast_fp16")];
+            tensor<int32, [4]> var_4808_perm_0 = const()[name = string("op_4808_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4809 = const()[name = string("op_4809"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_4808_cast_fp16 = transpose(perm = var_4808_perm_0, x = x_533_cast_fp16)[name = string("transpose_178")];
+            tensor<fp16, [1, 7, 1024]> input_1083_cast_fp16 = reshape(shape = var_4809, x = var_4808_cast_fp16)[name = string("input_1083_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418422208))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419470848))))[name = string("encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_20_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419472960)))];
+            tensor<fp16, [1, 7, 1024]> linear_187_cast_fp16 = linear(bias = encoder_layers_20_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_20_self_attn_linear_out_weight_to_fp16_quantized, x = input_1083_cast_fp16)[name = string("linear_187_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1087_cast_fp16 = add(x = input_1077_cast_fp16, y = linear_187_cast_fp16)[name = string("input_1087_cast_fp16")];
+            tensor<int32, [1]> x_537_axes_0 = const()[name = string("x_537_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419475072)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419477184)))];
+            tensor<fp16, [1, 7, 1024]> x_537_cast_fp16 = layer_norm(axes = x_537_axes_0, beta = encoder_layers_20_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_conv_weight_to_fp16, x = input_1087_cast_fp16)[name = string("x_537_cast_fp16")];
+            tensor<int32, [3]> input_1089_perm_0 = const()[name = string("input_1089_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1091_pad_type_0 = const()[name = string("input_1091_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1091_strides_0 = const()[name = string("input_1091_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1091_pad_0 = const()[name = string("input_1091_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1091_dilations_0 = const()[name = string("input_1091_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1091_groups_0 = const()[name = string("input_1091_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419479296))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421576512))))[name = string("encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_1089_cast_fp16 = transpose(perm = input_1089_perm_0, x = x_537_cast_fp16)[name = string("transpose_177")];
+            tensor<fp16, [1, 2048, 7]> input_1091_cast_fp16 = conv(dilations = input_1091_dilations_0, groups = input_1091_groups_0, pad = input_1091_pad_0, pad_type = input_1091_pad_type_0, strides = input_1091_strides_0, weight = encoder_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1089_cast_fp16)[name = string("input_1091_cast_fp16")];
+            int32 x_539_split_num_splits_0 = const()[name = string("x_539_split_num_splits_0"), val = int32(2)];
+            int32 x_539_split_axis_0 = const()[name = string("x_539_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_539_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_539_split_cast_fp16_1 = split(axis = x_539_split_axis_0, num_splits = x_539_split_num_splits_0, x = input_1091_cast_fp16)[name = string("x_539_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_539_split_1_sigmoid_cast_fp16 = sigmoid(x = x_539_split_cast_fp16_1)[name = string("x_539_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_539_cast_fp16 = mul(x = x_539_split_cast_fp16_0, y = x_539_split_1_sigmoid_cast_fp16)[name = string("x_539_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_1093_cast_fp16 = select(a = var_44_to_fp16, b = x_539_cast_fp16, cond = var_575)[name = string("input_1093_cast_fp16")];
+            bool new_x_83_interleave_0 = const()[name = string("new_x_83_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_83_cast_fp16 = concat(axis = var_59, interleave = new_x_83_interleave_0, values = (cache_83_cast_fp16, input_1093_cast_fp16))[name = string("new_x_83_cast_fp16")];
+            tensor<int32, [3]> var_4848_begin_0 = const()[name = string("op_4848_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_4848_end_0 = const()[name = string("op_4848_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_4848_end_mask_0 = const()[name = string("op_4848_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_4848_cast_fp16 = slice_by_index(begin = var_4848_begin_0, end = var_4848_end_0, end_mask = var_4848_end_mask_0, x = new_x_83_cast_fp16)[name = string("op_4848_cast_fp16")];
+            string x_541_pad_type_0 = const()[name = string("x_541_pad_type_0"), val = string("valid")];
+            int32 x_541_groups_0 = const()[name = string("x_541_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_541_strides_0 = const()[name = string("x_541_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_541_pad_0 = const()[name = string("x_541_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_541_dilations_0 = const()[name = string("x_541_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421580672))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421589952))))[name = string("encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_541_cast_fp16 = conv(dilations = x_541_dilations_0, groups = x_541_groups_0, pad = x_541_pad_0, pad_type = x_541_pad_type_0, strides = x_541_strides_0, weight = encoder_layers_20_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_83_cast_fp16)[name = string("x_541_cast_fp16")];
+            tensor<int32, [3]> input_1095_perm_0 = const()[name = string("input_1095_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_543_axes_0 = const()[name = string("x_543_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421592064)))];
+            tensor<fp16, [1024]> encoder_layers_20_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_20_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421594176)))];
+            tensor<fp16, [1, 7, 1024]> input_1095_cast_fp16 = transpose(perm = input_1095_perm_0, x = x_541_cast_fp16)[name = string("transpose_176")];
+            tensor<fp16, [1, 7, 1024]> x_543_cast_fp16 = layer_norm(axes = x_543_axes_0, beta = encoder_layers_20_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_conv_batch_norm_weight_to_fp16, x = input_1095_cast_fp16)[name = string("x_543_cast_fp16")];
+            tensor<int32, [3]> input_1097_perm_0 = const()[name = string("input_1097_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_1097_cast_fp16 = transpose(perm = input_1097_perm_0, x = x_543_cast_fp16)[name = string("transpose_175")];
+            tensor<fp16, [1, 1024, 7]> input_1099_cast_fp16 = silu(x = input_1097_cast_fp16)[name = string("input_1099_cast_fp16")];
+            string x_545_pad_type_0 = const()[name = string("x_545_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_545_strides_0 = const()[name = string("x_545_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_545_pad_0 = const()[name = string("x_545_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_545_dilations_0 = const()[name = string("x_545_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_545_groups_0 = const()[name = string("x_545_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421596288))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422644928))))[name = string("encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_545_cast_fp16 = conv(dilations = x_545_dilations_0, groups = x_545_groups_0, pad = x_545_pad_0, pad_type = x_545_pad_type_0, strides = x_545_strides_0, weight = encoder_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1099_cast_fp16)[name = string("x_545_cast_fp16")];
+            tensor<int32, [3]> input_1101_perm_0 = const()[name = string("input_1101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_1101_cast_fp16 = transpose(perm = input_1101_perm_0, x = x_545_cast_fp16)[name = string("transpose_174")];
+            tensor<fp16, [1, 7, 1024]> input_1103_cast_fp16 = add(x = input_1087_cast_fp16, y = input_1101_cast_fp16)[name = string("input_1103_cast_fp16")];
+            tensor<int32, [1]> input_1105_axes_0 = const()[name = string("input_1105_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422647040)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422649152)))];
+            tensor<fp16, [1, 7, 1024]> input_1105_cast_fp16 = layer_norm(axes = input_1105_axes_0, beta = encoder_layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_feed_forward2_weight_to_fp16, x = input_1103_cast_fp16)[name = string("input_1105_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(422651264))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426845632))))[name = string("encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_20_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426853888)))];
+            tensor<fp16, [1, 7, 4096]> linear_188_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1105_cast_fp16)[name = string("linear_188_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1109_cast_fp16 = silu(x = linear_188_cast_fp16)[name = string("input_1109_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426862144))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431056512))))[name = string("encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_20_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_20_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431058624)))];
+            tensor<fp16, [1, 7, 1024]> linear_189_cast_fp16 = linear(bias = encoder_layers_20_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_20_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1109_cast_fp16)[name = string("linear_189_cast_fp16")];
+            fp16 var_4891_to_fp16 = const()[name = string("op_4891_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4892_cast_fp16 = mul(x = linear_189_cast_fp16, y = var_4891_to_fp16)[name = string("op_4892_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1115_cast_fp16 = add(x = input_1103_cast_fp16, y = var_4892_cast_fp16)[name = string("input_1115_cast_fp16")];
+            tensor<int32, [1]> input_1117_axes_0 = const()[name = string("input_1117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_20_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_20_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431060736)))];
+            tensor<fp16, [1024]> encoder_layers_20_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_20_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431062848)))];
+            tensor<fp16, [1, 7, 1024]> input_1117_cast_fp16 = layer_norm(axes = input_1117_axes_0, beta = encoder_layers_20_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_20_norm_out_weight_to_fp16, x = input_1115_cast_fp16)[name = string("input_1117_cast_fp16")];
+            tensor<int32, [4]> cache_85_begin_0 = const()[name = string("cache_85_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
+            tensor<int32, [4]> cache_85_end_0 = const()[name = string("cache_85_end_0"), val = tensor<int32, [4]>([22, 1, 42, 1024])];
+            tensor<bool, [4]> cache_85_end_mask_0 = const()[name = string("cache_85_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_85_squeeze_mask_0 = const()[name = string("cache_85_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_85_cast_fp16 = slice_by_index(begin = cache_85_begin_0, end = cache_85_end_0, end_mask = cache_85_end_mask_0, squeeze_mask = cache_85_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_85_cast_fp16")];
+            tensor<int32, [4]> cache_87_begin_0 = const()[name = string("cache_87_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
+            tensor<int32, [4]> cache_87_end_0 = const()[name = string("cache_87_end_0"), val = tensor<int32, [4]>([22, 1, 1024, 8])];
+            tensor<bool, [4]> cache_87_end_mask_0 = const()[name = string("cache_87_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_87_squeeze_mask_0 = const()[name = string("cache_87_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_87_cast_fp16 = slice_by_index(begin = cache_87_begin_0, end = cache_87_end_0, end_mask = cache_87_end_mask_0, squeeze_mask = cache_87_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_87_cast_fp16")];
+            tensor<int32, [1]> input_1119_axes_0 = const()[name = string("input_1119_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431064960)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431067072)))];
+            tensor<fp16, [1, 7, 1024]> input_1119_cast_fp16 = layer_norm(axes = input_1119_axes_0, beta = encoder_layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_feed_forward1_weight_to_fp16, x = input_1117_cast_fp16)[name = string("input_1119_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431069184))), scale = tensor<fp16, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435263552))))[name = string("encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized")];
+            tensor<fp16, [4096]> encoder_layers_21_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435271808)))];
+            tensor<fp16, [1, 7, 4096]> linear_190_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1119_cast_fp16)[name = string("linear_190_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1123_cast_fp16 = silu(x = linear_190_cast_fp16)[name = string("input_1123_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435280064))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439474432))))[name = string("encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439476544)))];
+            tensor<fp16, [1, 7, 1024]> linear_191_cast_fp16 = linear(bias = encoder_layers_21_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1123_cast_fp16)[name = string("linear_191_cast_fp16")];
+            fp16 var_4928_to_fp16 = const()[name = string("op_4928_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_4929_cast_fp16 = mul(x = linear_191_cast_fp16, y = var_4928_to_fp16)[name = string("op_4929_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1129_cast_fp16 = add(x = input_1117_cast_fp16, y = var_4929_cast_fp16)[name = string("input_1129_cast_fp16")];
+            tensor<int32, [1]> key_43_axes_0 = const()[name = string("key_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439478656)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439480768)))];
+            tensor<fp16, [1, 7, 1024]> key_43_cast_fp16 = layer_norm(axes = key_43_axes_0, beta = encoder_layers_21_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_self_att_weight_to_fp16, x = input_1129_cast_fp16)[name = string("key_43_cast_fp16")];
+            bool input_1131_interleave_0 = const()[name = string("input_1131_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_1131_cast_fp16 = concat(axis = var_68, interleave = input_1131_interleave_0, values = (cache_85_cast_fp16, key_43_cast_fp16))[name = string("input_1131_cast_fp16")];
+            tensor<int32, [3]> var_4951_begin_0 = const()[name = string("op_4951_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_4951_end_0 = const()[name = string("op_4951_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_4951_end_mask_0 = const()[name = string("op_4951_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_4951_cast_fp16 = slice_by_index(begin = var_4951_begin_0, end = var_4951_end_0, end_mask = var_4951_end_mask_0, x = cache_85_cast_fp16)[name = string("op_4951_cast_fp16")];
+            bool var_4957_interleave_0 = const()[name = string("op_4957_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_4957_cast_fp16 = concat(axis = var_68, interleave = var_4957_interleave_0, values = (var_4951_cast_fp16, key_43_cast_fp16))[name = string("op_4957_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439482880))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440531520))))[name = string("encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440533632)))];
+            tensor<fp16, [1, 7, 1024]> linear_192_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_q_weight_to_fp16_quantized, x = key_43_cast_fp16)[name = string("linear_192_cast_fp16")];
+            tensor<int32, [4]> var_4962 = const()[name = string("op_4962"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_127_cast_fp16 = reshape(shape = var_4962, x = linear_192_cast_fp16)[name = string("q_127_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440535744))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441584384))))[name = string("encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441586496)))];
+            tensor<fp16, [1, 49, 1024]> linear_193_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_k_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_193_cast_fp16")];
+            tensor<int32, [4]> var_4967 = const()[name = string("op_4967"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_85_cast_fp16 = reshape(shape = var_4967, x = linear_193_cast_fp16)[name = string("k_85_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441588608))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442637248))))[name = string("encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized")];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442639360)))];
+            tensor<fp16, [1, 49, 1024]> linear_194_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_v_weight_to_fp16_quantized, x = input_1131_cast_fp16)[name = string("linear_194_cast_fp16")];
+            tensor<int32, [4]> var_4972 = const()[name = string("op_4972"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_43_cast_fp16 = reshape(shape = var_4972, x = linear_194_cast_fp16)[name = string("v_43_cast_fp16")];
+            tensor<int32, [4]> value_51_perm_0 = const()[name = string("value_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_21_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442641472)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4985_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_u_to_fp16)[name = string("op_4985_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_21_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_21_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442643584)))];
+            tensor<fp16, [1, 7, 8, 128]> var_4987_cast_fp16 = add(x = q_127_cast_fp16, y = encoder_layers_21_self_attn_pos_bias_v_to_fp16)[name = string("op_4987_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_43_perm_0 = const()[name = string("q_with_bias_v_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_553_transpose_x_0 = const()[name = string("x_553_transpose_x_0"), val = bool(false)];
+            bool x_553_transpose_y_0 = const()[name = string("x_553_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_4989_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442645696))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442745088))))[name = string("op_4989_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_43_cast_fp16 = transpose(perm = q_with_bias_v_43_perm_0, x = var_4987_cast_fp16)[name = string("transpose_173")];
+            tensor<fp16, [1, 8, 7, 97]> x_553_cast_fp16 = matmul(transpose_x = x_553_transpose_x_0, transpose_y = x_553_transpose_y_0, x = q_with_bias_v_43_cast_fp16, y = op_4989_to_fp16_quantized)[name = string("x_553_cast_fp16")];
+            tensor<int32, [8]> x_555_pad_0 = const()[name = string("x_555_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_555_mode_0 = const()[name = string("x_555_mode_0"), val = string("constant")];
+            fp16 const_352_to_fp16 = const()[name = string("const_352_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_555_cast_fp16 = pad(constant_val = const_352_to_fp16, mode = x_555_mode_0, pad = x_555_pad_0, x = x_553_cast_fp16)[name = string("x_555_cast_fp16")];
+            tensor<int32, [4]> var_4997 = const()[name = string("op_4997"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_557_cast_fp16 = reshape(shape = var_4997, x = x_555_cast_fp16)[name = string("x_557_cast_fp16")];
+            tensor<int32, [4]> var_5001_begin_0 = const()[name = string("op_5001_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5001_end_0 = const()[name = string("op_5001_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_5001_end_mask_0 = const()[name = string("op_5001_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_5001_cast_fp16 = slice_by_index(begin = var_5001_begin_0, end = var_5001_end_0, end_mask = var_5001_end_mask_0, x = x_557_cast_fp16)[name = string("op_5001_cast_fp16")];
+            tensor<int32, [4]> var_5002 = const()[name = string("op_5002"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_85_cast_fp16 = reshape(shape = var_5002, x = var_5001_cast_fp16)[name = string("matrix_bd_85_cast_fp16")];
+            bool matrix_ac_43_transpose_x_0 = const()[name = string("matrix_ac_43_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_43_transpose_y_0 = const()[name = string("matrix_ac_43_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_138_perm_0 = const()[name = string("transpose_138_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_139_perm_0 = const()[name = string("transpose_139_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_139 = transpose(perm = transpose_139_perm_0, x = k_85_cast_fp16)[name = string("transpose_171")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_138 = transpose(perm = transpose_138_perm_0, x = var_4985_cast_fp16)[name = string("transpose_172")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_43_cast_fp16 = matmul(transpose_x = matrix_ac_43_transpose_x_0, transpose_y = matrix_ac_43_transpose_y_0, x = transpose_138, y = transpose_139)[name = string("matrix_ac_43_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_87_begin_0 = const()[name = string("matrix_bd_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_87_end_0 = const()[name = string("matrix_bd_87_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_87_end_mask_0 = const()[name = string("matrix_bd_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_87_cast_fp16 = slice_by_index(begin = matrix_bd_87_begin_0, end = matrix_bd_87_end_0, end_mask = matrix_bd_87_end_mask_0, x = matrix_bd_85_cast_fp16)[name = string("matrix_bd_87_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_5011_cast_fp16 = add(x = matrix_ac_43_cast_fp16, y = matrix_bd_87_cast_fp16)[name = string("op_5011_cast_fp16")];
+            fp16 _inversed_scores_85_y_0_to_fp16 = const()[name = string("_inversed_scores_85_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_85_cast_fp16 = mul(x = var_5011_cast_fp16, y = _inversed_scores_85_y_0_to_fp16)[name = string("_inversed_scores_85_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_87_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_85_cast_fp16, cond = mask_11)[name = string("scores_87_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_5017_cast_fp16 = softmax(axis = var_59, x = scores_87_cast_fp16)[name = string("op_5017_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_1133_cast_fp16 = select(a = var_44_to_fp16, b = var_5017_cast_fp16, cond = mask_11)[name = string("input_1133_cast_fp16")];
+            bool x_559_transpose_x_0 = const()[name = string("x_559_transpose_x_0"), val = bool(false)];
+            bool x_559_transpose_y_0 = const()[name = string("x_559_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_51_cast_fp16 = transpose(perm = value_51_perm_0, x = v_43_cast_fp16)[name = string("transpose_170")];
+            tensor<fp16, [1, 8, 7, 128]> x_559_cast_fp16 = matmul(transpose_x = x_559_transpose_x_0, transpose_y = x_559_transpose_y_0, x = input_1133_cast_fp16, y = value_51_cast_fp16)[name = string("x_559_cast_fp16")];
+            tensor<int32, [4]> var_5021_perm_0 = const()[name = string("op_5021_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5022 = const()[name = string("op_5022"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_5021_cast_fp16 = transpose(perm = var_5021_perm_0, x = x_559_cast_fp16)[name = string("transpose_169")];
+            tensor<fp16, [1, 7, 1024]> input_1135_cast_fp16 = reshape(shape = var_5022, x = var_5021_cast_fp16)[name = string("input_1135_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_21_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(442745408)))];
+            tensor<fp16, [1024]> encoder_layers_21_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_21_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444842624)))];
+            tensor<fp16, [1, 7, 1024]> linear_196_cast_fp16 = linear(bias = encoder_layers_21_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_21_self_attn_linear_out_weight_to_fp16, x = input_1135_cast_fp16)[name = string("linear_196_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1139_cast_fp16 = add(x = input_1129_cast_fp16, y = linear_196_cast_fp16)[name = string("input_1139_cast_fp16")];
+            tensor<int32, [1]> x_563_axes_0 = const()[name = string("x_563_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444844736)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444846848)))];
+            tensor<fp16, [1, 7, 1024]> x_563_cast_fp16 = layer_norm(axes = x_563_axes_0, beta = encoder_layers_21_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_conv_weight_to_fp16, x = input_1139_cast_fp16)[name = string("x_563_cast_fp16")];
+            tensor<int32, [3]> input_1141_perm_0 = const()[name = string("input_1141_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1143_pad_type_0 = const()[name = string("input_1143_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1143_strides_0 = const()[name = string("input_1143_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1143_pad_0 = const()[name = string("input_1143_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1143_dilations_0 = const()[name = string("input_1143_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1143_groups_0 = const()[name = string("input_1143_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444848960))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446946176))))[name = string("encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_1141_cast_fp16 = transpose(perm = input_1141_perm_0, x = x_563_cast_fp16)[name = string("transpose_168")];
+            tensor<fp16, [1, 2048, 7]> input_1143_cast_fp16 = conv(dilations = input_1143_dilations_0, groups = input_1143_groups_0, pad = input_1143_pad_0, pad_type = input_1143_pad_type_0, strides = input_1143_strides_0, weight = encoder_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1141_cast_fp16)[name = string("input_1143_cast_fp16")];
+            int32 x_565_split_num_splits_0 = const()[name = string("x_565_split_num_splits_0"), val = int32(2)];
+            int32 x_565_split_axis_0 = const()[name = string("x_565_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_565_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_565_split_cast_fp16_1 = split(axis = x_565_split_axis_0, num_splits = x_565_split_num_splits_0, x = input_1143_cast_fp16)[name = string("x_565_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_565_split_1_sigmoid_cast_fp16 = sigmoid(x = x_565_split_cast_fp16_1)[name = string("x_565_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_565_cast_fp16 = mul(x = x_565_split_cast_fp16_0, y = x_565_split_1_sigmoid_cast_fp16)[name = string("x_565_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_1145_cast_fp16 = select(a = var_44_to_fp16, b = x_565_cast_fp16, cond = var_575)[name = string("input_1145_cast_fp16")];
+            bool new_x_87_interleave_0 = const()[name = string("new_x_87_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_87_cast_fp16 = concat(axis = var_59, interleave = new_x_87_interleave_0, values = (cache_87_cast_fp16, input_1145_cast_fp16))[name = string("new_x_87_cast_fp16")];
+            tensor<int32, [3]> var_5061_begin_0 = const()[name = string("op_5061_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_5061_end_0 = const()[name = string("op_5061_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_5061_end_mask_0 = const()[name = string("op_5061_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_5061_cast_fp16 = slice_by_index(begin = var_5061_begin_0, end = var_5061_end_0, end_mask = var_5061_end_mask_0, x = new_x_87_cast_fp16)[name = string("op_5061_cast_fp16")];
+            string x_567_pad_type_0 = const()[name = string("x_567_pad_type_0"), val = string("valid")];
+            int32 x_567_groups_0 = const()[name = string("x_567_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_567_strides_0 = const()[name = string("x_567_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_567_pad_0 = const()[name = string("x_567_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_567_dilations_0 = const()[name = string("x_567_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446950336))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446959616))))[name = string("encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_567_cast_fp16 = conv(dilations = x_567_dilations_0, groups = x_567_groups_0, pad = x_567_pad_0, pad_type = x_567_pad_type_0, strides = x_567_strides_0, weight = encoder_layers_21_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_87_cast_fp16)[name = string("x_567_cast_fp16")];
+            tensor<int32, [3]> input_1147_perm_0 = const()[name = string("input_1147_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_569_axes_0 = const()[name = string("x_569_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446961728)))];
+            tensor<fp16, [1024]> encoder_layers_21_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_21_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446963840)))];
+            tensor<fp16, [1, 7, 1024]> input_1147_cast_fp16 = transpose(perm = input_1147_perm_0, x = x_567_cast_fp16)[name = string("transpose_167")];
+            tensor<fp16, [1, 7, 1024]> x_569_cast_fp16 = layer_norm(axes = x_569_axes_0, beta = encoder_layers_21_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_conv_batch_norm_weight_to_fp16, x = input_1147_cast_fp16)[name = string("x_569_cast_fp16")];
+            tensor<int32, [3]> input_1149_perm_0 = const()[name = string("input_1149_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_1149_cast_fp16 = transpose(perm = input_1149_perm_0, x = x_569_cast_fp16)[name = string("transpose_166")];
+            tensor<fp16, [1, 1024, 7]> input_1151_cast_fp16 = silu(x = input_1149_cast_fp16)[name = string("input_1151_cast_fp16")];
+            string x_571_pad_type_0 = const()[name = string("x_571_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_571_strides_0 = const()[name = string("x_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_571_pad_0 = const()[name = string("x_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_571_dilations_0 = const()[name = string("x_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_571_groups_0 = const()[name = string("x_571_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446965952))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448014592))))[name = string("encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_571_cast_fp16 = conv(dilations = x_571_dilations_0, groups = x_571_groups_0, pad = x_571_pad_0, pad_type = x_571_pad_type_0, strides = x_571_strides_0, weight = encoder_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1151_cast_fp16)[name = string("x_571_cast_fp16")];
+            tensor<int32, [3]> input_1153_perm_0 = const()[name = string("input_1153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_1153_cast_fp16 = transpose(perm = input_1153_perm_0, x = x_571_cast_fp16)[name = string("transpose_165")];
+            tensor<fp16, [1, 7, 1024]> input_1155_cast_fp16 = add(x = input_1139_cast_fp16, y = input_1153_cast_fp16)[name = string("input_1155_cast_fp16")];
+            tensor<int32, [1]> input_1157_axes_0 = const()[name = string("input_1157_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448016704)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448018816)))];
+            tensor<fp16, [1, 7, 1024]> input_1157_cast_fp16 = layer_norm(axes = input_1157_axes_0, beta = encoder_layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_feed_forward2_weight_to_fp16, x = input_1155_cast_fp16)[name = string("input_1157_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_21_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(448020928)))];
+            tensor<fp16, [4096]> encoder_layers_21_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456409600)))];
+            tensor<fp16, [1, 7, 4096]> linear_197_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear1_weight_to_fp16, x = input_1157_cast_fp16)[name = string("linear_197_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1161_cast_fp16 = silu(x = linear_197_cast_fp16)[name = string("input_1161_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_21_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456417856)))];
+            tensor<fp16, [1024]> encoder_layers_21_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_21_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464806528)))];
+            tensor<fp16, [1, 7, 1024]> linear_198_cast_fp16 = linear(bias = encoder_layers_21_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_21_feed_forward2_linear2_weight_to_fp16, x = input_1161_cast_fp16)[name = string("linear_198_cast_fp16")];
+            fp16 var_5104_to_fp16 = const()[name = string("op_5104_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_5105_cast_fp16 = mul(x = linear_198_cast_fp16, y = var_5104_to_fp16)[name = string("op_5105_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1167_cast_fp16 = add(x = input_1155_cast_fp16, y = var_5105_cast_fp16)[name = string("input_1167_cast_fp16")];
+            tensor<int32, [1]> input_1169_axes_0 = const()[name = string("input_1169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_21_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_21_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464808640)))];
+            tensor<fp16, [1024]> encoder_layers_21_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_21_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464810752)))];
+            tensor<fp16, [1, 7, 1024]> input_1169_cast_fp16 = layer_norm(axes = input_1169_axes_0, beta = encoder_layers_21_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_21_norm_out_weight_to_fp16, x = input_1167_cast_fp16)[name = string("input_1169_cast_fp16")];
+            tensor<int32, [4]> cache_89_begin_0 = const()[name = string("cache_89_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
+            tensor<int32, [4]> cache_89_end_0 = const()[name = string("cache_89_end_0"), val = tensor<int32, [4]>([23, 1, 42, 1024])];
+            tensor<bool, [4]> cache_89_end_mask_0 = const()[name = string("cache_89_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_89_squeeze_mask_0 = const()[name = string("cache_89_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_89_cast_fp16 = slice_by_index(begin = cache_89_begin_0, end = cache_89_end_0, end_mask = cache_89_end_mask_0, squeeze_mask = cache_89_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_89_cast_fp16")];
+            tensor<int32, [4]> cache_91_begin_0 = const()[name = string("cache_91_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
+            tensor<int32, [4]> cache_91_end_0 = const()[name = string("cache_91_end_0"), val = tensor<int32, [4]>([23, 1, 1024, 8])];
+            tensor<bool, [4]> cache_91_end_mask_0 = const()[name = string("cache_91_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_91_squeeze_mask_0 = const()[name = string("cache_91_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_91_cast_fp16 = slice_by_index(begin = cache_91_begin_0, end = cache_91_end_0, end_mask = cache_91_end_mask_0, squeeze_mask = cache_91_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_91_cast_fp16")];
+            tensor<int32, [1]> input_1171_axes_0 = const()[name = string("input_1171_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464812864)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464814976)))];
+            tensor<fp16, [1, 7, 1024]> input_1171_cast_fp16 = layer_norm(axes = input_1171_axes_0, beta = encoder_layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_feed_forward1_weight_to_fp16, x = input_1169_cast_fp16)[name = string("input_1171_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_22_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464817088)))];
+            tensor<fp16, [4096]> encoder_layers_22_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473205760)))];
+            tensor<fp16, [1, 7, 4096]> linear_199_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear1_weight_to_fp16, x = input_1171_cast_fp16)[name = string("linear_199_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1175_cast_fp16 = silu(x = linear_199_cast_fp16)[name = string("input_1175_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_22_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473214016)))];
+            tensor<fp16, [1024]> encoder_layers_22_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481602688)))];
+            tensor<fp16, [1, 7, 1024]> linear_200_cast_fp16 = linear(bias = encoder_layers_22_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward1_linear2_weight_to_fp16, x = input_1175_cast_fp16)[name = string("linear_200_cast_fp16")];
+            fp16 var_5141_to_fp16 = const()[name = string("op_5141_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_5142_cast_fp16 = mul(x = linear_200_cast_fp16, y = var_5141_to_fp16)[name = string("op_5142_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1181_cast_fp16 = add(x = input_1169_cast_fp16, y = var_5142_cast_fp16)[name = string("input_1181_cast_fp16")];
+            tensor<int32, [1]> key_45_axes_0 = const()[name = string("key_45_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481604800)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481606912)))];
+            tensor<fp16, [1, 7, 1024]> key_45_cast_fp16 = layer_norm(axes = key_45_axes_0, beta = encoder_layers_22_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_self_att_weight_to_fp16, x = input_1181_cast_fp16)[name = string("key_45_cast_fp16")];
+            bool input_1183_interleave_0 = const()[name = string("input_1183_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_1183_cast_fp16 = concat(axis = var_68, interleave = input_1183_interleave_0, values = (cache_89_cast_fp16, key_45_cast_fp16))[name = string("input_1183_cast_fp16")];
+            tensor<int32, [3]> var_5164_begin_0 = const()[name = string("op_5164_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_5164_end_0 = const()[name = string("op_5164_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_5164_end_mask_0 = const()[name = string("op_5164_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_5164_cast_fp16 = slice_by_index(begin = var_5164_begin_0, end = var_5164_end_0, end_mask = var_5164_end_mask_0, x = cache_89_cast_fp16)[name = string("op_5164_cast_fp16")];
+            bool var_5170_interleave_0 = const()[name = string("op_5170_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> var_5170_cast_fp16 = concat(axis = var_68, interleave = var_5170_interleave_0, values = (var_5164_cast_fp16, key_45_cast_fp16))[name = string("op_5170_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481609024)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483706240)))];
+            tensor<fp16, [1, 7, 1024]> linear_201_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_q_weight_to_fp16, x = key_45_cast_fp16)[name = string("linear_201_cast_fp16")];
+            tensor<int32, [4]> var_5175 = const()[name = string("op_5175"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_133_cast_fp16 = reshape(shape = var_5175, x = linear_201_cast_fp16)[name = string("q_133_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483708352)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485805568)))];
+            tensor<fp16, [1, 49, 1024]> linear_202_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_k_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_202_cast_fp16")];
+            tensor<int32, [4]> var_5180 = const()[name = string("op_5180"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_89_cast_fp16 = reshape(shape = var_5180, x = linear_202_cast_fp16)[name = string("k_89_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485807680)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487904896)))];
+            tensor<fp16, [1, 49, 1024]> linear_203_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_v_weight_to_fp16, x = input_1183_cast_fp16)[name = string("linear_203_cast_fp16")];
+            tensor<int32, [4]> var_5185 = const()[name = string("op_5185"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_45_cast_fp16 = reshape(shape = var_5185, x = linear_203_cast_fp16)[name = string("v_45_cast_fp16")];
+            tensor<int32, [4]> value_53_perm_0 = const()[name = string("value_53_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_22_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487907008)))];
+            tensor<fp16, [1, 7, 8, 128]> var_5198_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_u_to_fp16)[name = string("op_5198_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_22_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_22_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487909120)))];
+            tensor<fp16, [1, 7, 8, 128]> var_5200_cast_fp16 = add(x = q_133_cast_fp16, y = encoder_layers_22_self_attn_pos_bias_v_to_fp16)[name = string("op_5200_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_45_perm_0 = const()[name = string("q_with_bias_v_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_579_transpose_x_0 = const()[name = string("x_579_transpose_x_0"), val = bool(false)];
+            bool x_579_transpose_y_0 = const()[name = string("x_579_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_5202_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487911232))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488010624))))[name = string("op_5202_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_45_cast_fp16 = transpose(perm = q_with_bias_v_45_perm_0, x = var_5200_cast_fp16)[name = string("transpose_164")];
+            tensor<fp16, [1, 8, 7, 97]> x_579_cast_fp16 = matmul(transpose_x = x_579_transpose_x_0, transpose_y = x_579_transpose_y_0, x = q_with_bias_v_45_cast_fp16, y = op_5202_to_fp16_quantized)[name = string("x_579_cast_fp16")];
+            tensor<int32, [8]> x_581_pad_0 = const()[name = string("x_581_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_581_mode_0 = const()[name = string("x_581_mode_0"), val = string("constant")];
+            fp16 const_365_to_fp16 = const()[name = string("const_365_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_581_cast_fp16 = pad(constant_val = const_365_to_fp16, mode = x_581_mode_0, pad = x_581_pad_0, x = x_579_cast_fp16)[name = string("x_581_cast_fp16")];
+            tensor<int32, [4]> var_5210 = const()[name = string("op_5210"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_583_cast_fp16 = reshape(shape = var_5210, x = x_581_cast_fp16)[name = string("x_583_cast_fp16")];
+            tensor<int32, [4]> var_5214_begin_0 = const()[name = string("op_5214_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5214_end_0 = const()[name = string("op_5214_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_5214_end_mask_0 = const()[name = string("op_5214_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_5214_cast_fp16 = slice_by_index(begin = var_5214_begin_0, end = var_5214_end_0, end_mask = var_5214_end_mask_0, x = x_583_cast_fp16)[name = string("op_5214_cast_fp16")];
+            tensor<int32, [4]> var_5215 = const()[name = string("op_5215"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_89_cast_fp16 = reshape(shape = var_5215, x = var_5214_cast_fp16)[name = string("matrix_bd_89_cast_fp16")];
+            bool matrix_ac_45_transpose_x_0 = const()[name = string("matrix_ac_45_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_45_transpose_y_0 = const()[name = string("matrix_ac_45_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_140_perm_0 = const()[name = string("transpose_140_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_141_perm_0 = const()[name = string("transpose_141_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_141 = transpose(perm = transpose_141_perm_0, x = k_89_cast_fp16)[name = string("transpose_162")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_140 = transpose(perm = transpose_140_perm_0, x = var_5198_cast_fp16)[name = string("transpose_163")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_45_cast_fp16 = matmul(transpose_x = matrix_ac_45_transpose_x_0, transpose_y = matrix_ac_45_transpose_y_0, x = transpose_140, y = transpose_141)[name = string("matrix_ac_45_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_91_begin_0 = const()[name = string("matrix_bd_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_91_end_0 = const()[name = string("matrix_bd_91_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_91_end_mask_0 = const()[name = string("matrix_bd_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_91_cast_fp16 = slice_by_index(begin = matrix_bd_91_begin_0, end = matrix_bd_91_end_0, end_mask = matrix_bd_91_end_mask_0, x = matrix_bd_89_cast_fp16)[name = string("matrix_bd_91_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_5224_cast_fp16 = add(x = matrix_ac_45_cast_fp16, y = matrix_bd_91_cast_fp16)[name = string("op_5224_cast_fp16")];
+            fp16 _inversed_scores_89_y_0_to_fp16 = const()[name = string("_inversed_scores_89_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_89_cast_fp16 = mul(x = var_5224_cast_fp16, y = _inversed_scores_89_y_0_to_fp16)[name = string("_inversed_scores_89_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_91_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_89_cast_fp16, cond = mask_11)[name = string("scores_91_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_5230_cast_fp16 = softmax(axis = var_59, x = scores_91_cast_fp16)[name = string("op_5230_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_1185_cast_fp16 = select(a = var_44_to_fp16, b = var_5230_cast_fp16, cond = mask_11)[name = string("input_1185_cast_fp16")];
+            bool x_585_transpose_x_0 = const()[name = string("x_585_transpose_x_0"), val = bool(false)];
+            bool x_585_transpose_y_0 = const()[name = string("x_585_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_53_cast_fp16 = transpose(perm = value_53_perm_0, x = v_45_cast_fp16)[name = string("transpose_161")];
+            tensor<fp16, [1, 8, 7, 128]> x_585_cast_fp16 = matmul(transpose_x = x_585_transpose_x_0, transpose_y = x_585_transpose_y_0, x = input_1185_cast_fp16, y = value_53_cast_fp16)[name = string("x_585_cast_fp16")];
+            tensor<int32, [4]> var_5234_perm_0 = const()[name = string("op_5234_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5235 = const()[name = string("op_5235"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_5234_cast_fp16 = transpose(perm = var_5234_perm_0, x = x_585_cast_fp16)[name = string("transpose_160")];
+            tensor<fp16, [1, 7, 1024]> input_1187_cast_fp16 = reshape(shape = var_5235, x = var_5234_cast_fp16)[name = string("input_1187_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_22_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488010944)))];
+            tensor<fp16, [1024]> encoder_layers_22_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_22_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490108160)))];
+            tensor<fp16, [1, 7, 1024]> linear_205_cast_fp16 = linear(bias = encoder_layers_22_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_22_self_attn_linear_out_weight_to_fp16, x = input_1187_cast_fp16)[name = string("linear_205_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1191_cast_fp16 = add(x = input_1181_cast_fp16, y = linear_205_cast_fp16)[name = string("input_1191_cast_fp16")];
+            tensor<int32, [1]> x_589_axes_0 = const()[name = string("x_589_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490110272)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490112384)))];
+            tensor<fp16, [1, 7, 1024]> x_589_cast_fp16 = layer_norm(axes = x_589_axes_0, beta = encoder_layers_22_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_conv_weight_to_fp16, x = input_1191_cast_fp16)[name = string("x_589_cast_fp16")];
+            tensor<int32, [3]> input_1193_perm_0 = const()[name = string("input_1193_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1195_pad_type_0 = const()[name = string("input_1195_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1195_strides_0 = const()[name = string("input_1195_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1195_pad_0 = const()[name = string("input_1195_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1195_dilations_0 = const()[name = string("input_1195_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1195_groups_0 = const()[name = string("input_1195_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490114496))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492211712))))[name = string("encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_1193_cast_fp16 = transpose(perm = input_1193_perm_0, x = x_589_cast_fp16)[name = string("transpose_159")];
+            tensor<fp16, [1, 2048, 7]> input_1195_cast_fp16 = conv(dilations = input_1195_dilations_0, groups = input_1195_groups_0, pad = input_1195_pad_0, pad_type = input_1195_pad_type_0, strides = input_1195_strides_0, weight = encoder_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1193_cast_fp16)[name = string("input_1195_cast_fp16")];
+            int32 x_591_split_num_splits_0 = const()[name = string("x_591_split_num_splits_0"), val = int32(2)];
+            int32 x_591_split_axis_0 = const()[name = string("x_591_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_591_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_591_split_cast_fp16_1 = split(axis = x_591_split_axis_0, num_splits = x_591_split_num_splits_0, x = input_1195_cast_fp16)[name = string("x_591_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_591_split_1_sigmoid_cast_fp16 = sigmoid(x = x_591_split_cast_fp16_1)[name = string("x_591_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_591_cast_fp16 = mul(x = x_591_split_cast_fp16_0, y = x_591_split_1_sigmoid_cast_fp16)[name = string("x_591_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_1197_cast_fp16 = select(a = var_44_to_fp16, b = x_591_cast_fp16, cond = var_575)[name = string("input_1197_cast_fp16")];
+            bool new_x_91_interleave_0 = const()[name = string("new_x_91_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_91_cast_fp16 = concat(axis = var_59, interleave = new_x_91_interleave_0, values = (cache_91_cast_fp16, input_1197_cast_fp16))[name = string("new_x_91_cast_fp16")];
+            tensor<int32, [3]> var_5274_begin_0 = const()[name = string("op_5274_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> var_5274_end_0 = const()[name = string("op_5274_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> var_5274_end_mask_0 = const()[name = string("op_5274_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> var_5274_cast_fp16 = slice_by_index(begin = var_5274_begin_0, end = var_5274_end_0, end_mask = var_5274_end_mask_0, x = new_x_91_cast_fp16)[name = string("op_5274_cast_fp16")];
+            string x_593_pad_type_0 = const()[name = string("x_593_pad_type_0"), val = string("valid")];
+            int32 x_593_groups_0 = const()[name = string("x_593_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_593_strides_0 = const()[name = string("x_593_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_593_pad_0 = const()[name = string("x_593_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_593_dilations_0 = const()[name = string("x_593_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492215872))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492225152))))[name = string("encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_593_cast_fp16 = conv(dilations = x_593_dilations_0, groups = x_593_groups_0, pad = x_593_pad_0, pad_type = x_593_pad_type_0, strides = x_593_strides_0, weight = encoder_layers_22_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_91_cast_fp16)[name = string("x_593_cast_fp16")];
+            tensor<int32, [3]> input_1199_perm_0 = const()[name = string("input_1199_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_595_axes_0 = const()[name = string("x_595_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492227264)))];
+            tensor<fp16, [1024]> encoder_layers_22_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_22_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492229376)))];
+            tensor<fp16, [1, 7, 1024]> input_1199_cast_fp16 = transpose(perm = input_1199_perm_0, x = x_593_cast_fp16)[name = string("transpose_158")];
+            tensor<fp16, [1, 7, 1024]> x_595_cast_fp16 = layer_norm(axes = x_595_axes_0, beta = encoder_layers_22_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_conv_batch_norm_weight_to_fp16, x = input_1199_cast_fp16)[name = string("x_595_cast_fp16")];
+            tensor<int32, [3]> input_1201_perm_0 = const()[name = string("input_1201_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_1201_cast_fp16 = transpose(perm = input_1201_perm_0, x = x_595_cast_fp16)[name = string("transpose_157")];
+            tensor<fp16, [1, 1024, 7]> input_1203_cast_fp16 = silu(x = input_1201_cast_fp16)[name = string("input_1203_cast_fp16")];
+            string x_597_pad_type_0 = const()[name = string("x_597_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_597_strides_0 = const()[name = string("x_597_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_597_pad_0 = const()[name = string("x_597_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_597_dilations_0 = const()[name = string("x_597_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_597_groups_0 = const()[name = string("x_597_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492231488))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493280128))))[name = string("encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_597_cast_fp16 = conv(dilations = x_597_dilations_0, groups = x_597_groups_0, pad = x_597_pad_0, pad_type = x_597_pad_type_0, strides = x_597_strides_0, weight = encoder_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1203_cast_fp16)[name = string("x_597_cast_fp16")];
+            tensor<int32, [3]> input_1205_perm_0 = const()[name = string("input_1205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_1205_cast_fp16 = transpose(perm = input_1205_perm_0, x = x_597_cast_fp16)[name = string("transpose_156")];
+            tensor<fp16, [1, 7, 1024]> input_1207_cast_fp16 = add(x = input_1191_cast_fp16, y = input_1205_cast_fp16)[name = string("input_1207_cast_fp16")];
+            tensor<int32, [1]> input_1209_axes_0 = const()[name = string("input_1209_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493282240)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493284352)))];
+            tensor<fp16, [1, 7, 1024]> input_1209_cast_fp16 = layer_norm(axes = input_1209_axes_0, beta = encoder_layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_feed_forward2_weight_to_fp16, x = input_1207_cast_fp16)[name = string("input_1209_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_22_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(493286464)))];
+            tensor<fp16, [4096]> encoder_layers_22_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501675136)))];
+            tensor<fp16, [1, 7, 4096]> linear_206_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear1_weight_to_fp16, x = input_1209_cast_fp16)[name = string("linear_206_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1213_cast_fp16 = silu(x = linear_206_cast_fp16)[name = string("input_1213_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_22_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501683392)))];
+            tensor<fp16, [1024]> encoder_layers_22_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_22_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510072064)))];
+            tensor<fp16, [1, 7, 1024]> linear_207_cast_fp16 = linear(bias = encoder_layers_22_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_22_feed_forward2_linear2_weight_to_fp16, x = input_1213_cast_fp16)[name = string("linear_207_cast_fp16")];
+            fp16 var_5317_to_fp16 = const()[name = string("op_5317_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_5318_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_5317_to_fp16)[name = string("op_5318_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1219_cast_fp16 = add(x = input_1207_cast_fp16, y = var_5318_cast_fp16)[name = string("input_1219_cast_fp16")];
+            tensor<int32, [1]> input_1221_axes_0 = const()[name = string("input_1221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_22_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_22_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510074176)))];
+            tensor<fp16, [1024]> encoder_layers_22_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_22_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510076288)))];
+            tensor<fp16, [1, 7, 1024]> input_1221_cast_fp16 = layer_norm(axes = input_1221_axes_0, beta = encoder_layers_22_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_22_norm_out_weight_to_fp16, x = input_1219_cast_fp16)[name = string("input_1221_cast_fp16")];
+            tensor<int32, [4]> cache_93_begin_0 = const()[name = string("cache_93_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
+            tensor<int32, [4]> cache_93_end_0 = const()[name = string("cache_93_end_0"), val = tensor<int32, [4]>([24, 1, 42, 1024])];
+            tensor<bool, [4]> cache_93_end_mask_0 = const()[name = string("cache_93_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_93_squeeze_mask_0 = const()[name = string("cache_93_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 42, 1024]> cache_93_cast_fp16 = slice_by_index(begin = cache_93_begin_0, end = cache_93_end_0, end_mask = cache_93_end_mask_0, squeeze_mask = cache_93_squeeze_mask_0, x = value_3_cast_fp16)[name = string("cache_93_cast_fp16")];
+            tensor<int32, [4]> cache_begin_0 = const()[name = string("cache_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
+            tensor<int32, [4]> cache_end_0 = const()[name = string("cache_end_0"), val = tensor<int32, [4]>([24, 1, 1024, 8])];
+            tensor<bool, [4]> cache_end_mask_0 = const()[name = string("cache_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_squeeze_mask_0 = const()[name = string("cache_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp16, [1, 1024, 8]> cache_cast_fp16 = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = value_5_cast_fp16)[name = string("cache_cast_fp16")];
+            tensor<int32, [1]> input_1223_axes_0 = const()[name = string("input_1223_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510078400)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510080512)))];
+            tensor<fp16, [1, 7, 1024]> input_1223_cast_fp16 = layer_norm(axes = input_1223_axes_0, beta = encoder_layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_feed_forward1_weight_to_fp16, x = input_1221_cast_fp16)[name = string("input_1223_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_23_feed_forward1_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510082624)))];
+            tensor<fp16, [4096]> encoder_layers_23_feed_forward1_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518471296)))];
+            tensor<fp16, [1, 7, 4096]> linear_208_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear1_weight_to_fp16, x = input_1223_cast_fp16)[name = string("linear_208_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1227_cast_fp16 = silu(x = linear_208_cast_fp16)[name = string("input_1227_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_23_feed_forward1_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518479552)))];
+            tensor<fp16, [1024]> encoder_layers_23_feed_forward1_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward1_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526868224)))];
+            tensor<fp16, [1, 7, 1024]> linear_209_cast_fp16 = linear(bias = encoder_layers_23_feed_forward1_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward1_linear2_weight_to_fp16, x = input_1227_cast_fp16)[name = string("linear_209_cast_fp16")];
+            fp16 var_5354_to_fp16 = const()[name = string("op_5354_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_5355_cast_fp16 = mul(x = linear_209_cast_fp16, y = var_5354_to_fp16)[name = string("op_5355_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1233_cast_fp16 = add(x = input_1221_cast_fp16, y = var_5355_cast_fp16)[name = string("input_1233_cast_fp16")];
+            tensor<int32, [1]> key_axes_0 = const()[name = string("key_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_self_att_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526870336)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_self_att_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526872448)))];
+            tensor<fp16, [1, 7, 1024]> key_cast_fp16 = layer_norm(axes = key_axes_0, beta = encoder_layers_23_norm_self_att_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_self_att_weight_to_fp16, x = input_1233_cast_fp16)[name = string("key_cast_fp16")];
+            bool input_1235_interleave_0 = const()[name = string("input_1235_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 49, 1024]> input_1235_cast_fp16 = concat(axis = var_68, interleave = input_1235_interleave_0, values = (cache_93_cast_fp16, key_cast_fp16))[name = string("input_1235_cast_fp16")];
+            tensor<int32, [3]> var_5377_begin_0 = const()[name = string("op_5377_begin_0"), val = tensor<int32, [3]>([0, 7, 0])];
+            tensor<int32, [3]> var_5377_end_0 = const()[name = string("op_5377_end_0"), val = tensor<int32, [3]>([1, 42, 1024])];
+            tensor<bool, [3]> var_5377_end_mask_0 = const()[name = string("op_5377_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 35, 1024]> var_5377_cast_fp16 = slice_by_index(begin = var_5377_begin_0, end = var_5377_end_0, end_mask = var_5377_end_mask_0, x = cache_93_cast_fp16)[name = string("op_5377_cast_fp16")];
+            bool cache_last_channel_cur_interleave_0 = const()[name = string("cache_last_channel_cur_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 42, 1024]> cache_last_channel_cur_cast_fp16 = concat(axis = var_68, interleave = cache_last_channel_cur_interleave_0, values = (var_5377_cast_fp16, key_cast_fp16))[name = string("cache_last_channel_cur_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_q_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526874560)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_q_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528971776)))];
+            tensor<fp16, [1, 7, 1024]> linear_210_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_q_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_q_weight_to_fp16, x = key_cast_fp16)[name = string("linear_210_cast_fp16")];
+            tensor<int32, [4]> var_5388 = const()[name = string("op_5388"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 7, 8, 128]> q_139_cast_fp16 = reshape(shape = var_5388, x = linear_210_cast_fp16)[name = string("q_139_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_k_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528973888)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_k_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531071104)))];
+            tensor<fp16, [1, 49, 1024]> linear_211_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_k_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_k_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_211_cast_fp16")];
+            tensor<int32, [4]> var_5393 = const()[name = string("op_5393"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> k_93_cast_fp16 = reshape(shape = var_5393, x = linear_211_cast_fp16)[name = string("k_93_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_v_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531073216)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_v_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533170432)))];
+            tensor<fp16, [1, 49, 1024]> linear_212_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_v_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_v_weight_to_fp16, x = input_1235_cast_fp16)[name = string("linear_212_cast_fp16")];
+            tensor<int32, [4]> var_5398 = const()[name = string("op_5398"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp16, [1, 49, 8, 128]> v_cast_fp16 = reshape(shape = var_5398, x = linear_212_cast_fp16)[name = string("v_cast_fp16")];
+            tensor<int32, [4]> value_perm_0 = const()[name = string("value_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp16, [8, 128]> encoder_layers_23_self_attn_pos_bias_u_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533172544)))];
+            tensor<fp16, [1, 7, 8, 128]> var_5411_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_u_to_fp16)[name = string("op_5411_cast_fp16")];
+            tensor<fp16, [8, 128]> encoder_layers_23_self_attn_pos_bias_v_to_fp16 = const()[name = string("encoder_layers_23_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533174656)))];
+            tensor<fp16, [1, 7, 8, 128]> var_5413_cast_fp16 = add(x = q_139_cast_fp16, y = encoder_layers_23_self_attn_pos_bias_v_to_fp16)[name = string("op_5413_cast_fp16")];
+            tensor<int32, [4]> q_with_bias_v_perm_0 = const()[name = string("q_with_bias_v_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool x_605_transpose_x_0 = const()[name = string("x_605_transpose_x_0"), val = bool(false)];
+            bool x_605_transpose_y_0 = const()[name = string("x_605_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 128, 97]> op_5415_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1, 8, 128, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533176768))), scale = tensor<fp16, [1, 1, 1, 97]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533276160))))[name = string("op_5415_to_fp16_quantized")];
+            tensor<fp16, [1, 8, 7, 128]> q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_5413_cast_fp16)[name = string("transpose_155")];
+            tensor<fp16, [1, 8, 7, 97]> x_605_cast_fp16 = matmul(transpose_x = x_605_transpose_x_0, transpose_y = x_605_transpose_y_0, x = q_with_bias_v_cast_fp16, y = op_5415_to_fp16_quantized)[name = string("x_605_cast_fp16")];
+            tensor<int32, [8]> x_607_pad_0 = const()[name = string("x_607_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            string x_607_mode_0 = const()[name = string("x_607_mode_0"), val = string("constant")];
+            fp16 const_378_to_fp16 = const()[name = string("const_378_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 8, 7, 98]> x_607_cast_fp16 = pad(constant_val = const_378_to_fp16, mode = x_607_mode_0, pad = x_607_pad_0, x = x_605_cast_fp16)[name = string("x_607_cast_fp16")];
+            tensor<int32, [4]> var_5423 = const()[name = string("op_5423"), val = tensor<int32, [4]>([1, 8, -1, 7])];
+            tensor<fp16, [1, 8, 98, 7]> x_609_cast_fp16 = reshape(shape = var_5423, x = x_607_cast_fp16)[name = string("x_609_cast_fp16")];
+            tensor<int32, [4]> var_5427_begin_0 = const()[name = string("op_5427_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5427_end_0 = const()[name = string("op_5427_end_0"), val = tensor<int32, [4]>([1, 8, 98, 7])];
+            tensor<bool, [4]> var_5427_end_mask_0 = const()[name = string("op_5427_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp16, [1, 8, 97, 7]> var_5427_cast_fp16 = slice_by_index(begin = var_5427_begin_0, end = var_5427_end_0, end_mask = var_5427_end_mask_0, x = x_609_cast_fp16)[name = string("op_5427_cast_fp16")];
+            tensor<int32, [4]> var_5428 = const()[name = string("op_5428"), val = tensor<int32, [4]>([1, 8, 7, 97])];
+            tensor<fp16, [1, 8, 7, 97]> matrix_bd_93_cast_fp16 = reshape(shape = var_5428, x = var_5427_cast_fp16)[name = string("matrix_bd_93_cast_fp16")];
+            bool matrix_ac_transpose_x_0 = const()[name = string("matrix_ac_transpose_x_0"), val = bool(false)];
+            bool matrix_ac_transpose_y_0 = const()[name = string("matrix_ac_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_142_perm_0 = const()[name = string("transpose_142_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_143_perm_0 = const()[name = string("transpose_143_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 128, 49]> transpose_143 = transpose(perm = transpose_143_perm_0, x = k_93_cast_fp16)[name = string("transpose_153")];
+            tensor<fp16, [1, 8, 7, 128]> transpose_142 = transpose(perm = transpose_142_perm_0, x = var_5411_cast_fp16)[name = string("transpose_154")];
+            tensor<fp16, [1, 8, 7, 49]> matrix_ac_cast_fp16 = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_142, y = transpose_143)[name = string("matrix_ac_cast_fp16")];
+            tensor<int32, [4]> matrix_bd_begin_0 = const()[name = string("matrix_bd_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_end_0 = const()[name = string("matrix_bd_end_0"), val = tensor<int32, [4]>([1, 8, 7, 49])];
+            tensor<bool, [4]> matrix_bd_end_mask_0 = const()[name = string("matrix_bd_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp16, [1, 8, 7, 49]> matrix_bd_cast_fp16 = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_93_cast_fp16)[name = string("matrix_bd_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_5437_cast_fp16 = add(x = matrix_ac_cast_fp16, y = matrix_bd_cast_fp16)[name = string("op_5437_cast_fp16")];
+            fp16 _inversed_scores_93_y_0_to_fp16 = const()[name = string("_inversed_scores_93_y_0_to_fp16"), val = fp16(0x1.6ap-4)];
+            tensor<fp16, [1, 8, 7, 49]> _inversed_scores_93_cast_fp16 = mul(x = var_5437_cast_fp16, y = _inversed_scores_93_y_0_to_fp16)[name = string("_inversed_scores_93_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> scores_cast_fp16 = select(a = var_45_to_fp16, b = _inversed_scores_93_cast_fp16, cond = mask_11)[name = string("scores_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> var_5443_cast_fp16 = softmax(axis = var_59, x = scores_cast_fp16)[name = string("op_5443_cast_fp16")];
+            tensor<fp16, [1, 8, 7, 49]> input_1237_cast_fp16 = select(a = var_44_to_fp16, b = var_5443_cast_fp16, cond = mask_11)[name = string("input_1237_cast_fp16")];
+            bool x_611_transpose_x_0 = const()[name = string("x_611_transpose_x_0"), val = bool(false)];
+            bool x_611_transpose_y_0 = const()[name = string("x_611_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 8, 49, 128]> value_cast_fp16 = transpose(perm = value_perm_0, x = v_cast_fp16)[name = string("transpose_152")];
+            tensor<fp16, [1, 8, 7, 128]> x_611_cast_fp16 = matmul(transpose_x = x_611_transpose_x_0, transpose_y = x_611_transpose_y_0, x = input_1237_cast_fp16, y = value_cast_fp16)[name = string("x_611_cast_fp16")];
+            tensor<int32, [4]> var_5447_perm_0 = const()[name = string("op_5447_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5448 = const()[name = string("op_5448"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp16, [1, 7, 8, 128]> var_5447_cast_fp16 = transpose(perm = var_5447_perm_0, x = x_611_cast_fp16)[name = string("transpose_151")];
+            tensor<fp16, [1, 7, 1024]> input_1239_cast_fp16 = reshape(shape = var_5448, x = var_5447_cast_fp16)[name = string("input_1239_cast_fp16")];
+            tensor<fp16, [1024, 1024]> encoder_layers_23_self_attn_linear_out_weight_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533276480)))];
+            tensor<fp16, [1024]> encoder_layers_23_self_attn_linear_out_bias_to_fp16 = const()[name = string("encoder_layers_23_self_attn_linear_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535373696)))];
+            tensor<fp16, [1, 7, 1024]> linear_214_cast_fp16 = linear(bias = encoder_layers_23_self_attn_linear_out_bias_to_fp16, weight = encoder_layers_23_self_attn_linear_out_weight_to_fp16, x = input_1239_cast_fp16)[name = string("linear_214_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1243_cast_fp16 = add(x = input_1233_cast_fp16, y = linear_214_cast_fp16)[name = string("input_1243_cast_fp16")];
+            tensor<int32, [1]> x_615_axes_0 = const()[name = string("x_615_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_conv_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535375808)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_conv_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535377920)))];
+            tensor<fp16, [1, 7, 1024]> x_615_cast_fp16 = layer_norm(axes = x_615_axes_0, beta = encoder_layers_23_norm_conv_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_conv_weight_to_fp16, x = input_1243_cast_fp16)[name = string("x_615_cast_fp16")];
+            tensor<int32, [3]> input_1245_perm_0 = const()[name = string("input_1245_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string input_1247_pad_type_0 = const()[name = string("input_1247_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> input_1247_strides_0 = const()[name = string("input_1247_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1247_pad_0 = const()[name = string("input_1247_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1247_dilations_0 = const()[name = string("input_1247_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 input_1247_groups_0 = const()[name = string("input_1247_groups_0"), val = int32(1)];
+            tensor<fp16, [2048, 1024, 1]> encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535380032))), scale = tensor<fp16, [2048, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537477248))))[name = string("encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> input_1245_cast_fp16 = transpose(perm = input_1245_perm_0, x = x_615_cast_fp16)[name = string("transpose_150")];
+            tensor<fp16, [1, 2048, 7]> input_1247_cast_fp16 = conv(dilations = input_1247_dilations_0, groups = input_1247_groups_0, pad = input_1247_pad_0, pad_type = input_1247_pad_type_0, strides = input_1247_strides_0, weight = encoder_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1245_cast_fp16)[name = string("input_1247_cast_fp16")];
+            int32 x_617_split_num_splits_0 = const()[name = string("x_617_split_num_splits_0"), val = int32(2)];
+            int32 x_617_split_axis_0 = const()[name = string("x_617_split_axis_0"), val = int32(1)];
+            tensor<fp16, [1, 1024, 7]> x_617_split_cast_fp16_0, tensor<fp16, [1, 1024, 7]> x_617_split_cast_fp16_1 = split(axis = x_617_split_axis_0, num_splits = x_617_split_num_splits_0, x = input_1247_cast_fp16)[name = string("x_617_split_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_617_split_1_sigmoid_cast_fp16 = sigmoid(x = x_617_split_cast_fp16_1)[name = string("x_617_split_1_sigmoid_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> x_617_cast_fp16 = mul(x = x_617_split_cast_fp16_0, y = x_617_split_1_sigmoid_cast_fp16)[name = string("x_617_cast_fp16")];
+            tensor<fp16, [1, 1024, 7]> input_1249_cast_fp16 = select(a = var_44_to_fp16, b = x_617_cast_fp16, cond = var_575)[name = string("input_1249_cast_fp16")];
+            bool new_x_interleave_0 = const()[name = string("new_x_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 1024, 15]> new_x_cast_fp16 = concat(axis = var_59, interleave = new_x_interleave_0, values = (cache_cast_fp16, input_1249_cast_fp16))[name = string("new_x_cast_fp16")];
+            tensor<int32, [3]> cache_last_time_cur_begin_0 = const()[name = string("cache_last_time_cur_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
+            tensor<int32, [3]> cache_last_time_cur_end_0 = const()[name = string("cache_last_time_cur_end_0"), val = tensor<int32, [3]>([1, 1024, 15])];
+            tensor<bool, [3]> cache_last_time_cur_end_mask_0 = const()[name = string("cache_last_time_cur_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 1024, 8]> cache_last_time_cur_cast_fp16 = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = new_x_cast_fp16)[name = string("cache_last_time_cur_cast_fp16")];
+            string x_619_pad_type_0 = const()[name = string("x_619_pad_type_0"), val = string("valid")];
+            int32 x_619_groups_0 = const()[name = string("x_619_groups_0"), val = int32(1024)];
+            tensor<int32, [1]> x_619_strides_0 = const()[name = string("x_619_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_619_pad_0 = const()[name = string("x_619_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_619_dilations_0 = const()[name = string("x_619_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1024, 1, 9]> encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537481408))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537490688))))[name = string("encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_619_cast_fp16 = conv(dilations = x_619_dilations_0, groups = x_619_groups_0, pad = x_619_pad_0, pad_type = x_619_pad_type_0, strides = x_619_strides_0, weight = encoder_layers_23_conv_depthwise_conv_weight_to_fp16_quantized, x = new_x_cast_fp16)[name = string("x_619_cast_fp16")];
+            tensor<int32, [3]> input_1251_perm_0 = const()[name = string("input_1251_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_621_axes_0 = const()[name = string("x_621_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_conv_batch_norm_weight_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537492800)))];
+            tensor<fp16, [1024]> encoder_layers_23_conv_batch_norm_bias_to_fp16 = const()[name = string("encoder_layers_23_conv_batch_norm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537494912)))];
+            tensor<fp16, [1, 7, 1024]> input_1251_cast_fp16 = transpose(perm = input_1251_perm_0, x = x_619_cast_fp16)[name = string("transpose_149")];
+            tensor<fp16, [1, 7, 1024]> x_621_cast_fp16 = layer_norm(axes = x_621_axes_0, beta = encoder_layers_23_conv_batch_norm_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_conv_batch_norm_weight_to_fp16, x = input_1251_cast_fp16)[name = string("x_621_cast_fp16")];
+            tensor<int32, [3]> input_1253_perm_0 = const()[name = string("input_1253_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 1024, 7]> input_1253_cast_fp16 = transpose(perm = input_1253_perm_0, x = x_621_cast_fp16)[name = string("transpose_148")];
+            tensor<fp16, [1, 1024, 7]> input_1255_cast_fp16 = silu(x = input_1253_cast_fp16)[name = string("input_1255_cast_fp16")];
+            string x_623_pad_type_0 = const()[name = string("x_623_pad_type_0"), val = string("valid")];
+            tensor<int32, [1]> x_623_strides_0 = const()[name = string("x_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_623_pad_0 = const()[name = string("x_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_623_dilations_0 = const()[name = string("x_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 x_623_groups_0 = const()[name = string("x_623_groups_0"), val = int32(1)];
+            tensor<fp16, [1024, 1024, 1]> encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537497024))), scale = tensor<fp16, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538545664))))[name = string("encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized")];
+            tensor<fp16, [1, 1024, 7]> x_623_cast_fp16 = conv(dilations = x_623_dilations_0, groups = x_623_groups_0, pad = x_623_pad_0, pad_type = x_623_pad_type_0, strides = x_623_strides_0, weight = encoder_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1255_cast_fp16)[name = string("x_623_cast_fp16")];
+            tensor<int32, [3]> input_1257_perm_0 = const()[name = string("input_1257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [1, 7, 1024]> input_1257_cast_fp16 = transpose(perm = input_1257_perm_0, x = x_623_cast_fp16)[name = string("transpose_147")];
+            tensor<fp16, [1, 7, 1024]> input_1259_cast_fp16 = add(x = input_1243_cast_fp16, y = input_1257_cast_fp16)[name = string("input_1259_cast_fp16")];
+            tensor<int32, [1]> input_1261_axes_0 = const()[name = string("input_1261_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538547776)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538549888)))];
+            tensor<fp16, [1, 7, 1024]> input_1261_cast_fp16 = layer_norm(axes = input_1261_axes_0, beta = encoder_layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_feed_forward2_weight_to_fp16, x = input_1259_cast_fp16)[name = string("input_1261_cast_fp16")];
+            tensor<fp16, [4096, 1024]> encoder_layers_23_feed_forward2_linear1_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538552000)))];
+            tensor<fp16, [4096]> encoder_layers_23_feed_forward2_linear1_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546940672)))];
+            tensor<fp16, [1, 7, 4096]> linear_215_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear1_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear1_weight_to_fp16, x = input_1261_cast_fp16)[name = string("linear_215_cast_fp16")];
+            tensor<fp16, [1, 7, 4096]> input_1265_cast_fp16 = silu(x = linear_215_cast_fp16)[name = string("input_1265_cast_fp16")];
+            tensor<fp16, [1024, 4096]> encoder_layers_23_feed_forward2_linear2_weight_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546948928)))];
+            tensor<fp16, [1024]> encoder_layers_23_feed_forward2_linear2_bias_to_fp16 = const()[name = string("encoder_layers_23_feed_forward2_linear2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555337600)))];
+            tensor<fp16, [1, 7, 1024]> linear_216_cast_fp16 = linear(bias = encoder_layers_23_feed_forward2_linear2_bias_to_fp16, weight = encoder_layers_23_feed_forward2_linear2_weight_to_fp16, x = input_1265_cast_fp16)[name = string("linear_216_cast_fp16")];
+            fp16 var_5530_to_fp16 = const()[name = string("op_5530_to_fp16"), val = fp16(0x1p-1)];
+            tensor<fp16, [1, 7, 1024]> var_5531_cast_fp16 = mul(x = linear_216_cast_fp16, y = var_5530_to_fp16)[name = string("op_5531_cast_fp16")];
+            tensor<fp16, [1, 7, 1024]> input_1271_cast_fp16 = add(x = input_1259_cast_fp16, y = var_5531_cast_fp16)[name = string("input_1271_cast_fp16")];
+            tensor<int32, [1]> audio_signal_axes_0 = const()[name = string("audio_signal_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [1024]> encoder_layers_23_norm_out_weight_to_fp16 = const()[name = string("encoder_layers_23_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555339712)))];
+            tensor<fp16, [1024]> encoder_layers_23_norm_out_bias_to_fp16 = const()[name = string("encoder_layers_23_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555341824)))];
+            tensor<fp16, [1, 7, 1024]> audio_signal_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = encoder_layers_23_norm_out_bias_to_fp16, epsilon = var_42_to_fp16, gamma = encoder_layers_23_norm_out_weight_to_fp16, x = input_1271_cast_fp16)[name = string("audio_signal_cast_fp16")];
+            int32 obj_5_axis_0 = const()[name = string("obj_5_axis_0"), val = int32(0)];
+            tensor<fp16, [24, 1, 42, 1024]> obj_5_cast_fp16 = stack(axis = obj_5_axis_0, values = (var_484_cast_fp16, var_697_cast_fp16, var_910_cast_fp16, var_1123_cast_fp16, var_1336_cast_fp16, var_1549_cast_fp16, var_1762_cast_fp16, var_1975_cast_fp16, var_2188_cast_fp16, var_2401_cast_fp16, var_2614_cast_fp16, var_2827_cast_fp16, var_3040_cast_fp16, var_3253_cast_fp16, var_3466_cast_fp16, var_3679_cast_fp16, var_3892_cast_fp16, var_4105_cast_fp16, var_4318_cast_fp16, var_4531_cast_fp16, var_4744_cast_fp16, var_4957_cast_fp16, var_5170_cast_fp16, cache_last_channel_cur_cast_fp16))[name = string("obj_5_cast_fp16")];
+            int32 obj_7_axis_0 = const()[name = string("obj_7_axis_0"), val = int32(0)];
+            tensor<fp16, [24, 1, 1024, 8]> obj_7_cast_fp16 = stack(axis = obj_7_axis_0, values = (var_588_cast_fp16, var_801_cast_fp16, var_1014_cast_fp16, var_1227_cast_fp16, var_1440_cast_fp16, var_1653_cast_fp16, var_1866_cast_fp16, var_2079_cast_fp16, var_2292_cast_fp16, var_2505_cast_fp16, var_2718_cast_fp16, var_2931_cast_fp16, var_3144_cast_fp16, var_3357_cast_fp16, var_3570_cast_fp16, var_3783_cast_fp16, var_3996_cast_fp16, var_4209_cast_fp16, var_4422_cast_fp16, var_4635_cast_fp16, var_4848_cast_fp16, var_5061_cast_fp16, var_5274_cast_fp16, cache_last_time_cur_cast_fp16))[name = string("obj_7_cast_fp16")];
+            tensor<int32, [1]> var_5547 = add(x = cache_len, y = max_audio_length_1)[name = string("op_5547")];
+            string var_5547_promoted_to_fp16_dtype_0 = const()[name = string("op_5547_promoted_to_fp16_dtype_0"), val = string("fp16")];
+            fp16 const_384_to_fp16 = const()[name = string("const_384_to_fp16"), val = fp16(-inf)];
+            fp16 var_49_promoted_to_fp16 = const()[name = string("op_49_promoted_to_fp16"), val = fp16(0x1.5p+5)];
+            tensor<fp16, [1]> var_5547_to_fp16 = cast(dtype = var_5547_promoted_to_fp16_dtype_0, x = var_5547)[name = string("cast_9")];
+            tensor<fp16, [1]> clip_1_cast_fp16 = clip(alpha = const_384_to_fp16, beta = var_49_promoted_to_fp16, x = var_5547_to_fp16)[name = string("clip_1_cast_fp16")];
+            int32 one_hot_1_batch_dims_0 = const()[name = string("one_hot_1_batch_dims_0"), val = int32(0)];
+            bool one_hot_1_validate_indices_0 = const()[name = string("one_hot_1_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [128, 128]> to_onehot_identity_table_to_fp16 = const()[name = string("to_onehot_identity_table_to_fp16"), val = tensor<fp16, [128, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555343936)))];
+            string prompt_id_to_int16_dtype_0 = const()[name = string("prompt_id_to_int16_dtype_0"), val = string("int16")];
+            string cast_230_dtype_0 = const()[name = string("cast_230_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
+            tensor<int16, [1]> prompt_id_to_int16 = cast(dtype = prompt_id_to_int16_dtype_0, x = prompt_id)[name = string("cast_8")];
+            tensor<int32, [1]> cast_230 = cast(dtype = cast_230_dtype_0, x = prompt_id_to_int16)[name = string("cast_7")];
+            tensor<bool, [1]> greater_equal_0 = greater_equal(x = cast_230, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
+            int32 slice_by_index_2 = const()[name = string("slice_by_index_2"), val = int32(128)];
+            tensor<int32, [1]> add_0 = add(x = cast_230, y = slice_by_index_2)[name = string("add_0")];
+            tensor<int32, [1]> select_0 = select(a = cast_230, b = add_0, cond = greater_equal_0)[name = string("select_0")];
+            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
+            string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
+            tensor<int16, [1]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_6")];
+            tensor<int32, [1]> cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_5")];
+            tensor<bool, [1]> greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
+            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(128)];
+            tensor<int32, [1]> add_0_1 = add(x = cast_0, y = slice_by_index_0)[name = string("add_0_1")];
+            tensor<int32, [1]> select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")];
+            int32 greater_equal_0_y_0_2 = const()[name = string("greater_equal_0_y_0_2"), val = int32(0)];
+            tensor<bool, [1]> greater_equal_0_2 = greater_equal(x = select_0_1, y = greater_equal_0_y_0_2)[name = string("greater_equal_0_2")];
+            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(128)];
+            tensor<int32, [1]> add_0_2 = add(x = select_0_1, y = slice_by_index_0_1)[name = string("add_0_2")];
+            tensor<int32, [1]> select_0_2 = select(a = select_0_1, b = add_0_2, cond = greater_equal_0_2)[name = string("select_0_2")];
+            int32 one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(0)];
+            tensor<fp16, [1, 128]> one_hot_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = one_hot_1_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = one_hot_1_batch_dims_0, indices = select_0_2, validate_indices = one_hot_1_validate_indices_0, x = to_onehot_identity_table_to_fp16)[name = string("one_hot_1_cast_fp16_cast_uint16_cast_uint16")];
+            tensor<int32, [1]> var_5593_axes_0 = const()[name = string("op_5593_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 128]> var_5593_cast_fp16 = expand_dims(axes = var_5593_axes_0, x = one_hot_1_cast_fp16_cast_uint16_cast_uint16)[name = string("op_5593_cast_fp16")];
+            tensor<int32, [3]> one_hot_reps_0 = const()[name = string("one_hot_reps_0"), val = tensor<int32, [3]>([1, 7, 1])];
+            tensor<fp16, [1, 7, 128]> one_hot_cast_fp16 = tile(reps = one_hot_reps_0, x = var_5593_cast_fp16)[name = string("one_hot_cast_fp16")];
+            int32 var_5602 = const()[name = string("op_5602"), val = int32(-1)];
+            bool input_1273_interleave_0 = const()[name = string("input_1273_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 7, 1152]> input_1273_cast_fp16 = concat(axis = var_5602, interleave = input_1273_interleave_0, values = (audio_signal_cast_fp16, one_hot_cast_fp16))[name = string("input_1273_cast_fp16")];
+            tensor<fp16, [2048, 1152]> prompt_kernel_0_weight_to_fp16 = const()[name = string("prompt_kernel_0_weight_to_fp16"), val = tensor<fp16, [2048, 1152]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555376768)))];
+            tensor<fp16, [2048]> prompt_kernel_0_bias_to_fp16 = const()[name = string("prompt_kernel_0_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560095424)))];
+            tensor<fp16, [1, 7, 2048]> linear_217_cast_fp16 = linear(bias = prompt_kernel_0_bias_to_fp16, weight = prompt_kernel_0_weight_to_fp16, x = input_1273_cast_fp16)[name = string("linear_217_cast_fp16")];
+            tensor<fp16, [1, 7, 2048]> input_cast_fp16 = relu(x = linear_217_cast_fp16)[name = string("input_cast_fp16")];
+            tensor<fp16, [1024, 2048]> prompt_kernel_2_weight_to_fp16 = const()[name = string("prompt_kernel_2_weight_to_fp16"), val = tensor<fp16, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560099584)))];
+            tensor<fp16, [1024]> prompt_kernel_2_bias_to_fp16 = const()[name = string("prompt_kernel_2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564293952)))];
+            tensor<fp16, [1, 7, 1024]> linear_218_cast_fp16 = linear(bias = prompt_kernel_2_bias_to_fp16, weight = prompt_kernel_2_weight_to_fp16, x = input_cast_fp16)[name = string("linear_218_cast_fp16")];
+            tensor<int32, [3]> var_5615_perm_0 = const()[name = string("op_5615_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string var_5615_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5615_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            string var_5620_dtype_0 = const()[name = string("op_5620_dtype_0"), val = string("int32")];
+            tensor<int32, [4]> var_5623_perm_0 = const()[name = string("op_5623_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string var_5623_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5623_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<int32, [4]> var_5626_perm_0 = const()[name = string("op_5626_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            string var_5626_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5626_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            string var_5631_dtype_0 = const()[name = string("op_5631_dtype_0"), val = string("int32")];
+            tensor<int32, [1]> cache_len_out = cast(dtype = var_5631_dtype_0, x = clip_1_cast_fp16)[name = string("cast_0")];
+            tensor<fp16, [1, 24, 1024, 8]> var_5626_cast_fp16 = transpose(perm = var_5626_perm_0, x = obj_7_cast_fp16)[name = string("transpose_144")];
+            tensor<fp32, [1, 24, 1024, 8]> cache_time_out = cast(dtype = var_5626_cast_fp16_to_fp32_dtype_0, x = var_5626_cast_fp16)[name = string("cast_1")];
+            tensor<fp16, [1, 24, 42, 1024]> var_5623_cast_fp16 = transpose(perm = var_5623_perm_0, x = obj_5_cast_fp16)[name = string("transpose_145")];
+            tensor<fp32, [1, 24, 42, 1024]> cache_channel_out = cast(dtype = var_5623_cast_fp16_to_fp32_dtype_0, x = var_5623_cast_fp16)[name = string("cast_2")];
+            tensor<int32, [1]> encoded_length = cast(dtype = var_5620_dtype_0, x = clip_0_cast_fp16)[name = string("cast_3")];
+            tensor<fp16, [1, 1024, 7]> var_5615_cast_fp16 = transpose(perm = var_5615_perm_0, x = linear_218_cast_fp16)[name = string("transpose_146")];
+            tensor<fp32, [1, 1024, 7]> encoded = cast(dtype = var_5615_cast_fp16_to_fp32_dtype_0, x = var_5615_cast_fp16)[name = string("cast_4")];
+        } -> (encoded, encoded_length, cache_channel_out, cache_time_out, cache_len_out);
+}
\ No newline at end of file
diff --git a/fr/560ms/encoder.mlmodelc/weights/weight.bin b/fr/560ms/encoder.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..7b1ca90fb42a5ee3e6703488fa2d418b42814a01
--- /dev/null
+++ b/fr/560ms/encoder.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:beaf958ec4cca237d94bf72f5afded9f1b3f6b93439b5d8e3b309e4ee1560e83
+size 564296064
diff --git a/fr/560ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/560ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..21020cc3a69303cc019d3c163864648766ab26b9
--- /dev/null
+++ b/fr/560ms/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e6d4ec3401a73a4b4c0409170717e2caf7bc829f85aa844dbb5f70309dc61ea
+size 802813
diff --git a/fr/560ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/560ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..7b1ca90fb42a5ee3e6703488fa2d418b42814a01
--- /dev/null
+++ b/fr/560ms/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:beaf958ec4cca237d94bf72f5afded9f1b3f6b93439b5d8e3b309e4ee1560e83
+size 564296064
diff --git a/fr/560ms/encoder.mlpackage/Manifest.json b/fr/560ms/encoder.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..44bfa973b64e3a625c9a52d15b370d5a8baa0bbd
--- /dev/null
+++ b/fr/560ms/encoder.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "5BB04C05-9A8D-40F1-9561-33AB8C6304FC": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "7CFD7E55-2E07-4224-889C-214DD954BA80": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "7CFD7E55-2E07-4224-889C-214DD954BA80"
+}
diff --git a/fr/560ms/joint.mlmodelc/analytics/coremldata.bin b/fr/560ms/joint.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..38aade026cc9376471863f9f09be292c9fe8578b
--- /dev/null
+++ b/fr/560ms/joint.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad1f6f5571ee9e7ad8a533b4679ed7416879f4d8e5fc10f40835c1d014b822a4
+size 243
diff --git a/fr/560ms/joint.mlmodelc/coremldata.bin b/fr/560ms/joint.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..45df627e7953068f36ea8f3e5467664cf318a44e
--- /dev/null
+++ b/fr/560ms/joint.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e630dfefcd0f652f2815f76cfbacb89b51f2c94618fbc517e41779426099a1f
+size 341
diff --git a/fr/560ms/joint.mlmodelc/model.mil b/fr/560ms/joint.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..041dd87f66df1673f6226d5c642c93a64d4e2730
--- /dev/null
+++ b/fr/560ms/joint.mlmodelc/model.mil
@@ -0,0 +1,31 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios18>(tensor<fp32, [1, 640, 1]> decoder, tensor<fp32, [1, 1024, 1]> encoder) {
+            tensor<int32, [3]> input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string encoder_to_fp16_dtype_0 = const()[name = string("encoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [640, 1024]> module_enc_weight_to_fp16 = const()[name = string("module_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [640]> module_enc_bias_to_fp16 = const()[name = string("module_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1310848)))];
+            tensor<fp16, [1, 1024, 1]> encoder_to_fp16 = cast(dtype = encoder_to_fp16_dtype_0, x = encoder)[name = string("cast_2")];
+            tensor<fp16, [1, 1, 1024]> input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = encoder_to_fp16)[name = string("transpose_1")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = module_enc_bias_to_fp16, weight = module_enc_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> module_pred_weight_to_fp16 = const()[name = string("module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1312192)))];
+            tensor<fp16, [640]> module_pred_bias_to_fp16 = const()[name = string("module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2131456)))];
+            tensor<fp16, [1, 640, 1]> decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_1")];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = decoder_to_fp16)[name = string("transpose_0")];
+            tensor<fp16, [1, 1, 640]> linear_1_cast_fp16 = linear(bias = module_pred_bias_to_fp16, weight = module_pred_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_23_axes_0 = const()[name = string("op_23_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_23_cast_fp16 = expand_dims(axes = var_23_axes_0, x = linear_0_cast_fp16)[name = string("op_23_cast_fp16")];
+            tensor<int32, [1]> var_25_axes_0 = const()[name = string("op_25_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_25_cast_fp16 = expand_dims(axes = var_25_axes_0, x = linear_1_cast_fp16)[name = string("op_25_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_5_cast_fp16 = add(x = var_23_cast_fp16, y = var_25_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<fp16, [1, 1, 1, 640]> input_7_cast_fp16 = relu(x = input_5_cast_fp16)[name = string("input_7_cast_fp16")];
+            tensor<fp16, [849, 640]> module_joint_net_2_weight_to_fp16 = const()[name = string("module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2132800)))];
+            tensor<fp16, [849]> module_joint_net_2_bias_to_fp16 = const()[name = string("module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [849]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3219584)))];
+            tensor<fp16, [1, 1, 1, 849]> linear_2_cast_fp16 = linear(bias = module_joint_net_2_bias_to_fp16, weight = module_joint_net_2_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_2_cast_fp16")];
+            string linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_2_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 1, 1, 849]> logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_0")];
+        } -> (logits);
+}
\ No newline at end of file
diff --git a/fr/560ms/joint.mlmodelc/weights/weight.bin b/fr/560ms/joint.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..109a5bf4eb27d11fcd040ea350a8e6d00d18932a
--- /dev/null
+++ b/fr/560ms/joint.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9fa5fb0d14d87484d9ffd1f4d30996ea456c14ea3e5ab9d910486d8e86d3a540
+size 3221346
diff --git a/fr/560ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/560ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..0c650fe487d18d0d937a1fc295a3eee53b4aacbe
--- /dev/null
+++ b/fr/560ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4bac39a8a21b3e7e8cdeba1d03b92e8f743bf87cb2de3b1d85c505a580a0b218
+size 4486
diff --git a/fr/560ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/560ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..109a5bf4eb27d11fcd040ea350a8e6d00d18932a
--- /dev/null
+++ b/fr/560ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9fa5fb0d14d87484d9ffd1f4d30996ea456c14ea3e5ab9d910486d8e86d3a540
+size 3221346
diff --git a/fr/560ms/joint.mlpackage/Manifest.json b/fr/560ms/joint.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7e49a7390ba8e327aff397061f0fd4476b7c70a
--- /dev/null
+++ b/fr/560ms/joint.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "0DC18C6B-B3DE-4D3A-80EE-28CAA06F6708": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "88105E4A-1C49-40F4-A835-C9B725CEEC59": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "88105E4A-1C49-40F4-A835-C9B725CEEC59"
+}
diff --git a/fr/560ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin b/fr/560ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..a856ad184cad37664eb30bb699b3a88e6401000d
--- /dev/null
+++ b/fr/560ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28b5c13f8ce2d8154edb31968c174d322ee4ae76db42f41ecf7ac5e0fd87f57d
+size 243
diff --git a/fr/560ms/joint_noencproj_batched.mlmodelc/coremldata.bin b/fr/560ms/joint_noencproj_batched.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..b68dc2ebdbe0b8957cd45ed4b265fda6b6195204
--- /dev/null
+++ b/fr/560ms/joint_noencproj_batched.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5dc74778713432d34163ae80901d3294eef3341dd020a29dc05d67eba8f6f0a
+size 406
diff --git a/fr/560ms/joint_noencproj_batched.mlmodelc/model.mil b/fr/560ms/joint_noencproj_batched.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..b66479ae4eaa92652a98a34a18e993259a14e6c0
--- /dev/null
+++ b/fr/560ms/joint_noencproj_batched.mlmodelc/model.mil
@@ -0,0 +1,26 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios18>(tensor<fp32, [1, 640, 1]> decoder, tensor<fp32, [1, 4, 640]> encoder_proj) {
+            tensor<int32, [3]> input_1_perm_0 = const()[name = string("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            string decoder_to_fp16_dtype_0 = const()[name = string("decoder_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [640, 640]> joint_module_pred_weight_to_fp16 = const()[name = string("joint_module_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [640]> joint_module_pred_bias_to_fp16 = const()[name = string("joint_module_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819328)))];
+            tensor<fp16, [1, 640, 1]> decoder_to_fp16 = cast(dtype = decoder_to_fp16_dtype_0, x = decoder)[name = string("cast_2")];
+            tensor<fp16, [1, 1, 640]> input_1_cast_fp16 = transpose(perm = input_1_perm_0, x = decoder_to_fp16)[name = string("transpose_0")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_module_pred_bias_to_fp16, weight = joint_module_pred_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [1]> var_15_axes_0 = const()[name = string("op_15_axes_0"), val = tensor<int32, [1]>([2])];
+            string encoder_proj_to_fp16_dtype_0 = const()[name = string("encoder_proj_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 4, 640]> encoder_proj_to_fp16 = cast(dtype = encoder_proj_to_fp16_dtype_0, x = encoder_proj)[name = string("cast_1")];
+            tensor<fp16, [1, 4, 1, 640]> var_15_cast_fp16 = expand_dims(axes = var_15_axes_0, x = encoder_proj_to_fp16)[name = string("op_15_cast_fp16")];
+            tensor<int32, [1]> var_17_axes_0 = const()[name = string("op_17_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1, 640]> var_17_cast_fp16 = expand_dims(axes = var_17_axes_0, x = linear_0_cast_fp16)[name = string("op_17_cast_fp16")];
+            tensor<fp16, [1, 4, 1, 640]> input_3_cast_fp16 = add(x = var_15_cast_fp16, y = var_17_cast_fp16)[name = string("input_3_cast_fp16")];
+            tensor<fp16, [1, 4, 1, 640]> input_5_cast_fp16 = relu(x = input_3_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<fp16, [849, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [849, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(820672)))];
+            tensor<fp16, [849]> joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [849]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1907456)))];
+            tensor<fp16, [1, 4, 1, 849]> linear_1_cast_fp16 = linear(bias = joint_module_joint_net_2_bias_to_fp16, weight = joint_module_joint_net_2_weight_to_fp16, x = input_5_cast_fp16)[name = string("linear_1_cast_fp16")];
+            string linear_1_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_1_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 4, 1, 849]> logits = cast(dtype = linear_1_cast_fp16_to_fp32_dtype_0, x = linear_1_cast_fp16)[name = string("cast_0")];
+        } -> (logits);
+}
\ No newline at end of file
diff --git a/fr/560ms/joint_noencproj_batched.mlmodelc/weights/weight.bin b/fr/560ms/joint_noencproj_batched.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..4e66049a9a8bde2b92859df3192b671ddf9ab81d
--- /dev/null
+++ b/fr/560ms/joint_noencproj_batched.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9a9dd24e6b63b97540b0db0e990418bc6fddc9c2bf85b82eb226e6df1ea02a4
+size 1909218
diff --git a/fr/560ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/560ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..4f82db74749dc3350042c3f6c1453c4d4e620d50
--- /dev/null
+++ b/fr/560ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d13f08c4961187e47001773d7d4d5d5937ba7ff68475eb7c1bac41d7014c64a1
+size 3839
diff --git a/fr/560ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/560ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..4e66049a9a8bde2b92859df3192b671ddf9ab81d
--- /dev/null
+++ b/fr/560ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9a9dd24e6b63b97540b0db0e990418bc6fddc9c2bf85b82eb226e6df1ea02a4
+size 1909218
diff --git a/fr/560ms/joint_noencproj_batched.mlpackage/Manifest.json b/fr/560ms/joint_noencproj_batched.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..5f2e3080d9f6dc5617b9630128af453e31562411
--- /dev/null
+++ b/fr/560ms/joint_noencproj_batched.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "EFDDA9F1-A088-4CE2-8A64-5928EB8F429C": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "F28FF7B3-AC0A-4ACA-93F8-DF201DFFE75B": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "F28FF7B3-AC0A-4ACA-93F8-DF201DFFE75B"
+}
diff --git a/fr/560ms/metadata.json b/fr/560ms/metadata.json
new file mode 100644
index 0000000000000000000000000000000000000000..eca67dd447986a332cf7121403407165387fc78c
--- /dev/null
+++ b/fr/560ms/metadata.json
@@ -0,0 +1,199 @@
+{
+  "model": "nvidia/nemotron-asr-streaming-multilingual-0.6b",
+  "model_class": "nemo.collections.asr.models.rnnt_bpe_models_prompt.EncDecRNNTBPEModelWithPrompt",
+  "sample_rate": 16000,
+  "mel_features": 128,
+  "chunk_mel_frames": 56,
+  "pre_encode_cache": 9,
+  "total_mel_frames": 65,
+  "att_context_size": [
+    42,
+    13
+  ],
+  "vocab_size": 848,
+  "blank_idx": 848,
+  "vocab_pruned": true,
+  "vocab_pruned_original_size": 13087,
+  "cache_channel_shape": [
+    1,
+    24,
+    42,
+    1024
+  ],
+  "cache_time_shape": [
+    1,
+    24,
+    1024,
+    8
+  ],
+  "decoder_hidden": 640,
+  "decoder_layers": 2,
+  "encoder_dim": 1024,
+  "num_prompts": 128,
+  "prompt_dictionary": {
+    "af-ZA": 54,
+    "am-ET": 49,
+    "ar": 7,
+    "ar-AR": 7,
+    "auto": 101,
+    "ay-BO": 81,
+    "az-AZ": 66,
+    "bg": 30,
+    "bg-BG": 30,
+    "bn-IN": 36,
+    "cs": 22,
+    "cs-CZ": 22,
+    "da": 25,
+    "da-DK": 25,
+    "de": 9,
+    "de-DE": 9,
+    "el": 21,
+    "el-GR": 21,
+    "en": 0,
+    "en-GB": 1,
+    "en-US": 0,
+    "enGB": 1,
+    "es": 3,
+    "es-ES": 2,
+    "es-US": 3,
+    "esES": 2,
+    "et": 60,
+    "et-EE": 60,
+    "fa-IR": 38,
+    "fi": 26,
+    "fi-FI": 26,
+    "fr": 8,
+    "fr-CA": 100,
+    "fr-FR": 8,
+    "gn-PY": 82,
+    "gu-IN": 42,
+    "ha-NG": 50,
+    "haw-US": 97,
+    "he-IL": 64,
+    "hi": 6,
+    "hi-HI": 6,
+    "hi-IN": 6,
+    "hr": 29,
+    "hr-HR": 29,
+    "hu": 23,
+    "hu-HU": 23,
+    "hy-AM": 68,
+    "id-ID": 34,
+    "ig-NG": 53,
+    "it": 15,
+    "it-IT": 15,
+    "ja-JA": 10,
+    "ja-JP": 10,
+    "ka-GE": 67,
+    "km-KH": 47,
+    "kn-IN": 43,
+    "ko": 14,
+    "ko-KO": 14,
+    "ko-KR": 14,
+    "ku-TR": 65,
+    "ky-KG": 71,
+    "ln-CD": 58,
+    "lt": 31,
+    "lt-LT": 31,
+    "lv": 61,
+    "lv-LV": 61,
+    "mi-NZ": 96,
+    "ml-IN": 44,
+    "mr-IN": 41,
+    "ms-MY": 35,
+    "mt-MT": 102,
+    "nah-MX": 83,
+    "nb": 103,
+    "nb-NO": 103,
+    "ne-NP": 46,
+    "nl": 16,
+    "nl-NL": 16,
+    "nn": 104,
+    "nn-NO": 104,
+    "no": 27,
+    "no-NO": 27,
+    "ny-MW": 57,
+    "or-KE": 59,
+    "pl": 17,
+    "pl-PL": 17,
+    "pt": 13,
+    "pt-BR": 12,
+    "pt-PT": 13,
+    "qu-PE": 80,
+    "ro": 20,
+    "ro-RO": 20,
+    "ru": 11,
+    "ru-RU": 11,
+    "rw-RW": 55,
+    "si-LK": 45,
+    "sk": 28,
+    "sk-SK": 28,
+    "sl": 62,
+    "sl-SI": 62,
+    "sm-WS": 98,
+    "so-SO": 56,
+    "sv": 24,
+    "sv-SE": 24,
+    "sw-KE": 48,
+    "ta-IN": 39,
+    "te-IN": 40,
+    "tg-TJ": 70,
+    "th-TH": 32,
+    "to-TO": 99,
+    "tr": 18,
+    "tr-TR": 18,
+    "uk": 19,
+    "uk-UA": 19,
+    "ur-PK": 37,
+    "uz-UZ": 69,
+    "vi-VN": 33,
+    "yo-NG": 52,
+    "zh-CN": 4,
+    "zh-TW": 5,
+    "zh-ZH": 4,
+    "zu-ZA": 51
+  },
+  "default_prompt_id": 101,
+  "lang_tag_token_ids": [
+    1,
+    72,
+    122,
+    170,
+    230,
+    231,
+    263,
+    282,
+    390,
+    415,
+    430,
+    483,
+    501,
+    519,
+    542,
+    565,
+    594,
+    619,
+    620,
+    629,
+    640,
+    659,
+    660,
+    677,
+    704,
+    717,
+    735,
+    741,
+    790,
+    791,
+    793,
+    794,
+    795,
+    800,
+    819,
+    829,
+    830,
+    844,
+    845
+  ],
+  "chunk_ms": 560
+}
\ No newline at end of file
diff --git a/fr/560ms/preprocessor.mlmodelc/analytics/coremldata.bin b/fr/560ms/preprocessor.mlmodelc/analytics/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..d4b1561ab413a9d87db506bc842f077779dcbded
--- /dev/null
+++ b/fr/560ms/preprocessor.mlmodelc/analytics/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b8033bec5ee01649f325b8f4c5aeef1b31c99b469ce56d46039c1b73f09585d
+size 243
diff --git a/fr/560ms/preprocessor.mlmodelc/coremldata.bin b/fr/560ms/preprocessor.mlmodelc/coremldata.bin
new file mode 100644
index 0000000000000000000000000000000000000000..50c6f5fdff7aca33e5a111a3ad36751334f6bcbc
--- /dev/null
+++ b/fr/560ms/preprocessor.mlmodelc/coremldata.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da7084cfdd3b4e9c7c91f3fac1a37afcce7ae905e8b799e72f99aa4fdf6664ab
+size 431
diff --git a/fr/560ms/preprocessor.mlmodelc/model.mil b/fr/560ms/preprocessor.mlmodelc/model.mil
new file mode 100644
index 0000000000000000000000000000000000000000..b1a0b2b9193c992de42e51245fc1ef433d345afc
--- /dev/null
+++ b/fr/560ms/preprocessor.mlmodelc/model.mil
@@ -0,0 +1,122 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.10.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios18>(tensor<fp32, [1, ?]> audio, tensor<int32, [1]> audio_length) [FlexibleShapeInformation = tuple<tuple<string, dict<string, tensor<int32, [?]>>>, tuple<string, dict<string, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"audio", [1, 1]}}), ("RangeDims", {{"audio", [[1, 1], [1, 1280000]]}})))] {
+            int32 var_9 = const()[name = string("op_9"), val = int32(1)];
+            int32 var_10 = const()[name = string("op_10"), val = int32(160)];
+            int32 var_12 = const()[name = string("op_12"), val = int32(0)];
+            int32 var_33 = const()[name = string("op_33"), val = int32(512)];
+            tensor<int32, [1]> var_34 = add(x = audio_length, y = var_33)[name = string("op_34")];
+            int32 var_35 = const()[name = string("op_35"), val = int32(512)];
+            tensor<int32, [1]> var_36 = sub(x = var_34, y = var_35)[name = string("op_36")];
+            tensor<int32, [1]> floor_div_0 = floor_div(x = var_36, y = var_10)[name = string("floor_div_0")];
+            tensor<bool, [1]> var_39 = equal(x = audio_length, y = var_12)[name = string("op_39")];
+            tensor<int32, [1]> var_40 = const()[name = string("op_40"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> mel_length = select(a = var_40, b = floor_div_0, cond = var_39)[name = string("seq_len")];
+            string audio_to_fp16_dtype_0 = const()[name = string("audio_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, ?]> audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = string("cast_10")];
+            tensor<int32, [2]> var_42_shape_cast_fp16 = shape(x = audio_to_fp16)[name = string("op_42_shape_cast_fp16")];
+            int32 gather_0_axis_0 = const()[name = string("gather_0_axis_0"), val = int32(0)];
+            int32 gather_0_batch_dims_0 = const()[name = string("gather_0_batch_dims_0"), val = int32(0)];
+            bool gather_0_validate_indices_0 = const()[name = string("gather_0_validate_indices_0"), val = bool(false)];
+            string var_42_shape_cast_fp16_to_int16_dtype_0 = const()[name = string("op_42_shape_cast_fp16_to_int16_dtype_0"), val = string("int16")];
+            uint16 gather_0_indices_0_to_uint16 = const()[name = string("gather_0_indices_0_to_uint16"), val = uint16(1)];
+            tensor<int16, [2]> var_42_shape_cast_fp16_to_int16 = cast(dtype = var_42_shape_cast_fp16_to_int16_dtype_0, x = var_42_shape_cast_fp16)[name = string("cast_9")];
+            int16 gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = gather_0_indices_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_42_shape_cast_fp16_to_int16)[name = string("gather_0_cast_uint16")];
+            string gather_0_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_0_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            int32 const_0 = const()[name = string("const_0"), val = int32(0)];
+            int32 const_1 = const()[name = string("const_1"), val = int32(1)];
+            int32 gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = string("cast_8")];
+            tensor<int32, [?]> var_43 = range_1d(end = gather_0_cast_uint16_to_int32, start = const_0, step = const_1)[name = string("op_43")];
+            tensor<int32, [1]> var_44_axes_0 = const()[name = string("op_44_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> var_44 = expand_dims(axes = var_44_axes_0, x = var_43)[name = string("op_44")];
+            tensor<int32, [1]> var_45_axes_0 = const()[name = string("op_45_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_45 = expand_dims(axes = var_45_axes_0, x = audio_length)[name = string("op_45")];
+            tensor<bool, [1, ?]> timemask = less(x = var_44, y = var_45)[name = string("timemask")];
+            tensor<int32, [2]> var_48_begin_0 = const()[name = string("op_48_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_48_end_0 = const()[name = string("op_48_end_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<bool, [2]> var_48_end_mask_0 = const()[name = string("op_48_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<bool, [2]> var_48_squeeze_mask_0 = const()[name = string("op_48_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
+            tensor<fp16, [1]> var_48_cast_fp16 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, squeeze_mask = var_48_squeeze_mask_0, x = audio_to_fp16)[name = string("op_48_cast_fp16")];
+            tensor<int32, [1]> var_49_axes_0 = const()[name = string("op_49_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1]> var_49_cast_fp16 = expand_dims(axes = var_49_axes_0, x = var_48_cast_fp16)[name = string("op_49_cast_fp16")];
+            tensor<int32, [2]> var_51_begin_0 = const()[name = string("op_51_begin_0"), val = tensor<int32, [2]>([0, 1])];
+            tensor<int32, [2]> var_51_end_0 = const()[name = string("op_51_end_0"), val = tensor<int32, [2]>([1, 0])];
+            tensor<bool, [2]> var_51_end_mask_0 = const()[name = string("op_51_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<fp16, [1, ?]> var_51_cast_fp16 = slice_by_index(begin = var_51_begin_0, end = var_51_end_0, end_mask = var_51_end_mask_0, x = audio_to_fp16)[name = string("op_51_cast_fp16")];
+            tensor<int32, [2]> var_53_begin_0 = const()[name = string("op_53_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_53_end_0 = const()[name = string("op_53_end_0"), val = tensor<int32, [2]>([1, -1])];
+            tensor<bool, [2]> var_53_end_mask_0 = const()[name = string("op_53_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp16, [1, ?]> var_53_cast_fp16 = slice_by_index(begin = var_53_begin_0, end = var_53_end_0, end_mask = var_53_end_mask_0, x = audio_to_fp16)[name = string("op_53_cast_fp16")];
+            fp16 var_54_to_fp16 = const()[name = string("op_54_to_fp16"), val = fp16(0x1.f0cp-1)];
+            tensor<fp16, [1, ?]> var_55_cast_fp16 = mul(x = var_53_cast_fp16, y = var_54_to_fp16)[name = string("op_55_cast_fp16")];
+            tensor<fp16, [1, ?]> var_56_cast_fp16 = sub(x = var_51_cast_fp16, y = var_55_cast_fp16)[name = string("op_56_cast_fp16")];
+            bool x_3_interleave_0 = const()[name = string("x_3_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, ?]> x_3_cast_fp16 = concat(axis = var_9, interleave = x_3_interleave_0, values = (var_49_cast_fp16, var_56_cast_fp16))[name = string("x_3_cast_fp16")];
+            tensor<bool, [1, ?]> var_59 = logical_not(x = timemask)[name = string("op_59")];
+            fp16 var_16_to_fp16 = const()[name = string("op_16_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, ?]> input_1_cast_fp16 = select(a = var_16_to_fp16, b = x_3_cast_fp16, cond = var_59)[name = string("input_1_cast_fp16")];
+            tensor<int32, [3]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp16, [1, 1, ?]> input_3_cast_fp16 = reshape(shape = concat_1x, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")];
+            tensor<int32, [6]> input_5_pad_0 = const()[name = string("input_5_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 256, 256])];
+            string input_5_mode_0 = const()[name = string("input_5_mode_0"), val = string("constant")];
+            fp16 const_3_to_fp16 = const()[name = string("const_3_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, ?]> input_5_cast_fp16 = pad(constant_val = const_3_to_fp16, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")];
+            tensor<int32, [2]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [2]>([1, -1])];
+            tensor<fp16, [1, ?]> input_cast_fp16 = reshape(shape = concat_2x, x = input_5_cast_fp16)[name = string("input_cast_fp16")];
+            tensor<int32, [1]> expand_dims_3 = const()[name = string("expand_dims_3"), val = tensor<int32, [1]>([160])];
+            tensor<int32, [1]> expand_dims_4_axes_0 = const()[name = string("expand_dims_4_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, ?]> expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = input_cast_fp16)[name = string("expand_dims_4_cast_fp16")];
+            string conv_0_pad_type_0 = const()[name = string("conv_0_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> conv_0_pad_0 = const()[name = string("conv_0_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_0_dilations_0 = const()[name = string("conv_0_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 conv_0_groups_0 = const()[name = string("conv_0_groups_0"), val = int32(1)];
+            tensor<fp16, [257, 1, 512]> expand_dims_1_to_fp16 = const()[name = string("expand_dims_1_to_fp16"), val = tensor<fp16, [257, 1, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 257, ?]> conv_0_cast_fp16 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_0_cast_fp16")];
+            string conv_1_pad_type_0 = const()[name = string("conv_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> conv_1_pad_0 = const()[name = string("conv_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_1_dilations_0 = const()[name = string("conv_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            int32 conv_1_groups_0 = const()[name = string("conv_1_groups_0"), val = int32(1)];
+            tensor<fp16, [257, 1, 512]> expand_dims_2_to_fp16 = const()[name = string("expand_dims_2_to_fp16"), val = tensor<fp16, [257, 1, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(263296)))];
+            tensor<fp16, [1, 257, ?]> conv_1_cast_fp16 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2_to_fp16, x = expand_dims_4_cast_fp16)[name = string("conv_1_cast_fp16")];
+            int32 stack_0_axis_0 = const()[name = string("stack_0_axis_0"), val = int32(-1)];
+            tensor<fp16, [1, 257, ?, 2]> stack_0_cast_fp16 = stack(axis = stack_0_axis_0, values = (conv_0_cast_fp16, conv_1_cast_fp16))[name = string("stack_0_cast_fp16")];
+            fp16 var_19_promoted_to_fp16 = const()[name = string("op_19_promoted_to_fp16"), val = fp16(0x1p+1)];
+            tensor<fp16, [1, 257, ?, 2]> var_74_cast_fp16 = pow(x = stack_0_cast_fp16, y = var_19_promoted_to_fp16)[name = string("op_74_cast_fp16")];
+            tensor<int32, [1]> var_76_axes_0 = const()[name = string("op_76_axes_0"), val = tensor<int32, [1]>([-1])];
+            bool var_76_keep_dims_0 = const()[name = string("op_76_keep_dims_0"), val = bool(false)];
+            tensor<fp16, [1, 257, ?]> var_76_cast_fp16 = reduce_sum(axes = var_76_axes_0, keep_dims = var_76_keep_dims_0, x = var_74_cast_fp16)[name = string("op_76_cast_fp16")];
+            tensor<fp16, [1, 257, ?]> x_11_cast_fp16 = identity(x = var_76_cast_fp16)[name = string("x_11_cast_fp16")];
+            bool x_13_transpose_x_0 = const()[name = string("x_13_transpose_x_0"), val = bool(false)];
+            bool x_13_transpose_y_0 = const()[name = string("x_13_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 128, 257]> const_4_to_fp16 = const()[name = string("const_4_to_fp16"), val = tensor<fp16, [1, 128, 257]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(526528)))];
+            tensor<fp16, [1, 128, ?]> x_13_cast_fp16 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = const_4_to_fp16, y = x_11_cast_fp16)[name = string("x_13_cast_fp16")];
+            fp16 var_83_to_fp16 = const()[name = string("op_83_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 128, ?]> var_84_cast_fp16 = add(x = x_13_cast_fp16, y = var_83_to_fp16)[name = string("op_84_cast_fp16")];
+            fp32 x_epsilon_0 = const()[name = string("x_epsilon_0"), val = fp32(0x1p-149)];
+            tensor<fp16, [1, 128, ?]> x_cast_fp16 = log(epsilon = x_epsilon_0, x = var_84_cast_fp16)[name = string("x_cast_fp16")];
+            tensor<int32, [3]> var_86_shape_cast_fp16 = shape(x = x_cast_fp16)[name = string("op_86_shape_cast_fp16")];
+            int32 gather_5_batch_dims_0 = const()[name = string("gather_5_batch_dims_0"), val = int32(0)];
+            bool gather_5_validate_indices_0 = const()[name = string("gather_5_validate_indices_0"), val = bool(false)];
+            string var_86_shape_cast_fp16_to_uint16_dtype_0 = const()[name = string("op_86_shape_cast_fp16_to_uint16_dtype_0"), val = string("uint16")];
+            int32 gather_5_cast_uint16_axis_0 = const()[name = string("gather_5_cast_uint16_axis_0"), val = int32(0)];
+            uint16 select_0_to_uint16 = const()[name = string("select_0_to_uint16"), val = uint16(2)];
+            tensor<uint16, [3]> var_86_shape_cast_fp16_to_uint16 = cast(dtype = var_86_shape_cast_fp16_to_uint16_dtype_0, x = var_86_shape_cast_fp16)[name = string("cast_7")];
+            uint16 gather_5_cast_uint16_cast_uint16 = gather(axis = gather_5_cast_uint16_axis_0, batch_dims = gather_5_batch_dims_0, indices = select_0_to_uint16, validate_indices = gather_5_validate_indices_0, x = var_86_shape_cast_fp16_to_uint16)[name = string("gather_5_cast_uint16_cast_uint16")];
+            string gather_5_cast_uint16_to_int32_dtype_0 = const()[name = string("gather_5_cast_uint16_to_int32_dtype_0"), val = string("int32")];
+            int32 const_5 = const()[name = string("const_5"), val = int32(0)];
+            int32 const_6 = const()[name = string("const_6"), val = int32(1)];
+            int32 gather_5_cast_uint16_to_int32 = cast(dtype = gather_5_cast_uint16_to_int32_dtype_0, x = gather_5_cast_uint16_cast_uint16)[name = string("cast_6")];
+            tensor<int32, [?]> mask_1 = range_1d(end = gather_5_cast_uint16_to_int32, start = const_5, step = const_6)[name = string("mask_1")];
+            tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = string("expand_dims_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = string("expand_dims_0")];
+            tensor<int32, [1]> var_91_axes_0 = const()[name = string("op_91_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_91 = expand_dims(axes = var_91_axes_0, x = mel_length)[name = string("op_91")];
+            tensor<bool, [1, ?]> mask = greater_equal(x = expand_dims_0, y = var_91)[name = string("mask")];
+            tensor<int32, [1]> var_93_axes_0 = const()[name = string("op_93_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, ?]> var_93 = expand_dims(axes = var_93_axes_0, x = mask)[name = string("op_93")];
+            tensor<fp16, [1, 128, ?]> processed_signal_cast_fp16 = select(a = var_16_to_fp16, b = x_cast_fp16, cond = var_93)[name = string("processed_signal_cast_fp16")];
+            string processed_signal_cast_fp16_to_fp32_dtype_0 = const()[name = string("processed_signal_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 128, ?]> mel = cast(dtype = processed_signal_cast_fp16_to_fp32_dtype_0, x = processed_signal_cast_fp16)[name = string("cast_5")];
+        } -> (mel, mel_length);
+}
\ No newline at end of file
diff --git a/fr/560ms/preprocessor.mlmodelc/weights/weight.bin b/fr/560ms/preprocessor.mlmodelc/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411
--- /dev/null
+++ b/fr/560ms/preprocessor.mlmodelc/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3
+size 592384
diff --git a/fr/560ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel b/fr/560ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel
new file mode 100644
index 0000000000000000000000000000000000000000..050fa97ca7a2aa4b7c4fa318f4fa2a51914287c4
--- /dev/null
+++ b/fr/560ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0feb90f2cf9a5bc749fa763b6eb78a23755b6184b41a262fe08faebe4e709b3e
+size 16035
diff --git a/fr/560ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/fr/560ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin
new file mode 100644
index 0000000000000000000000000000000000000000..86dd375f6649d262d58c9cf8b89006ceab216411
--- /dev/null
+++ b/fr/560ms/preprocessor.mlpackage/Data/com.apple.CoreML/weights/weight.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3
+size 592384
diff --git a/fr/560ms/preprocessor.mlpackage/Manifest.json b/fr/560ms/preprocessor.mlpackage/Manifest.json
new file mode 100644
index 0000000000000000000000000000000000000000..11f4407dd2c2a1fafa63dd69a93da5d4e2866fba
--- /dev/null
+++ b/fr/560ms/preprocessor.mlpackage/Manifest.json
@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "571B03EE-25FB-448D-B106-8C0685101326": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "B1E3045B-7172-4628-BD63-F2960CEFBA16": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "571B03EE-25FB-448D-B106-8C0685101326"
+}
diff --git a/fr/560ms/tokenizer.json b/fr/560ms/tokenizer.json
new file mode 100644
index 0000000000000000000000000000000000000000..53c1afe82c2a72808c78fffb846345d590c453fd
--- /dev/null
+++ b/fr/560ms/tokenizer.json
@@ -0,0 +1,851 @@
+{
+  "0": "<unk>",
+  "1": "<bg-BG>",
+  "2": "▁",
+  "3": ".",
+  "4": ",",
+  "5": "e",
+  "6": "t",
+  "7": "a",
+  "8": "s",
+  "9": "o",
+  "10": "i",
+  "11": "r",
+  "12": "l",
+  "13": "u",
+  "14": "d",
+  "15": "c",
+  "16": "h",
+  "17": "m",
+  "18": "p",
+  "19": "n",
+  "20": "g",
+  "21": "f",
+  "22": "en",
+  "23": "in",
+  "24": "on",
+  "25": "y",
+  "26": "er",
+  "27": "an",
+  "28": "w",
+  "29": "0",
+  "30": "b",
+  "31": "v",
+  "32": "2",
+  "33": "1",
+  "34": "▁the",
+  "35": "k",
+  "36": "3",
+  "37": "5",
+  "38": "q",
+  "39": "I",
+  "40": "é",
+  "41": "4",
+  "42": "z",
+  "43": "A",
+  "44": "6",
+  "45": "j",
+  "46": "E",
+  "47": "7",
+  "48": "T",
+  "49": "8",
+  "50": "9",
+  "51": "S",
+  "52": "x",
+  "53": "C",
+  "54": "M",
+  "55": "P",
+  "56": "D",
+  "57": "B",
+  "58": "U",
+  "59": "N",
+  "60": "O",
+  "61": "G",
+  "62": "à",
+  "63": "L",
+  "64": "è",
+  "65": "V",
+  "66": "R",
+  "67": "ó",
+  "68": "J",
+  "69": "K",
+  "70": "ê",
+  "71": "ç",
+  "72": "<cs-CZ>",
+  "73": "st",
+  "74": "ch",
+  "75": "le",
+  "76": "li",
+  "77": "▁po",
+  "78": "no",
+  "79": "to",
+  "80": "me",
+  "81": "te",
+  "82": "ho",
+  "83": "▁pro",
+  "84": "ro",
+  "85": "▁na",
+  "86": "ce",
+  "87": "la",
+  "88": "ni",
+  "89": "ra",
+  "90": "ti",
+  "91": "lo",
+  "92": "ko",
+  "93": "po",
+  "94": "je",
+  "95": "de",
+  "96": "na",
+  "97": "mi",
+  "98": "ci",
+  "99": "▁by",
+  "100": "ve",
+  "101": "▁za",
+  "102": "▁A",
+  "103": "re",
+  "104": "ou",
+  "105": "vo",
+  "106": "né",
+  "107": "va",
+  "108": "mo",
+  "109": "ze",
+  "110": "ne",
+  "111": "ka",
+  "112": "▁jak",
+  "113": "ny",
+  "114": "vi",
+  "115": "ent",
+  "116": "▁pan",
+  "117": "Z",
+  "118": "▁To",
+  "119": "▁Je",
+  "120": "ă",
+  "121": "X",
+  "122": "<da-DK>",
+  "123": "▁for",
+  "124": "▁at",
+  "125": "et",
+  "126": "▁vi",
+  "127": "al",
+  "128": "▁de",
+  "129": "▁der",
+  "130": "or",
+  "131": "om",
+  "132": "and",
+  "133": "at",
+  "134": "▁af",
+  "135": "ge",
+  "136": "ar",
+  "137": "is",
+  "138": "▁be",
+  "139": "un",
+  "140": "lig",
+  "141": "▁man",
+  "142": "ig",
+  "143": "▁som",
+  "144": "el",
+  "145": "ag",
+  "146": "erne",
+  "147": "▁den",
+  "148": "ste",
+  "149": "id",
+  "150": "ske",
+  "151": "iv",
+  "152": "▁her",
+  "153": "ion",
+  "154": "am",
+  "155": "ur",
+  "156": "for",
+  "157": "▁pr",
+  "158": "▁sig",
+  "159": "▁ind",
+  "160": "ende",
+  "161": "▁Vi",
+  "162": "ation",
+  "163": "mme",
+  "164": "tage",
+  "165": "▁op",
+  "166": "▁fra",
+  "167": "▁alle",
+  "168": "▁Men",
+  "169": "▁var",
+  "170": "<de-DE>",
+  "171": "▁die",
+  "172": "sch",
+  "173": "it",
+  "174": "gen",
+  "175": "▁W",
+  "176": "▁B",
+  "177": "▁E",
+  "178": "▁F",
+  "179": "ll",
+  "180": "▁es",
+  "181": "▁K",
+  "182": "ie",
+  "183": "au",
+  "184": "ich",
+  "185": "ck",
+  "186": "ten",
+  "187": "mal",
+  "188": "ein",
+  "189": "▁T",
+  "190": "▁mit",
+  "191": "hr",
+  "192": "ter",
+  "193": "tz",
+  "194": "▁G",
+  "195": "ben",
+  "196": "um",
+  "197": "us",
+  "198": "cht",
+  "199": "il",
+  "200": "ut",
+  "201": "▁ver",
+  "202": "▁hier",
+  "203": "▁von",
+  "204": "ri",
+  "205": "ach",
+  "206": "ol",
+  "207": "▁Da",
+  "208": "sp",
+  "209": "ell",
+  "210": "▁ja",
+  "211": "uch",
+  "212": "wi",
+  "213": "rei",
+  "214": "▁Ge",
+  "215": "und",
+  "216": "isch",
+  "217": "▁sie",
+  "218": "▁Ja",
+  "219": "▁du",
+  "220": "▁dem",
+  "221": "▁Sch",
+  "222": "▁Ma",
+  "223": "▁De",
+  "224": "▁Le",
+  "225": "▁In",
+  "226": "▁Re",
+  "227": "▁sein",
+  "228": "▁Mi",
+  "229": "▁Ha",
+  "230": "<el-GR>",
+  "231": "<et-EE>",
+  "232": "ma",
+  "233": "ta",
+  "234": "se",
+  "235": "da",
+  "236": "si",
+  "237": "ks",
+  "238": "ga",
+  "239": "he",
+  "240": "mu",
+  "241": "tu",
+  "242": "ha",
+  "243": "ja",
+  "244": "gi",
+  "245": "nd",
+  "246": "oo",
+  "247": "gu",
+  "248": "ju",
+  "249": "est",
+  "250": "▁pa",
+  "251": "tud",
+  "252": "▁pea",
+  "253": "sta",
+  "254": "di",
+  "255": "mise",
+  "256": "▁su",
+  "257": "ide",
+  "258": "val",
+  "259": "▁Me",
+  "260": "ment",
+  "261": "▁meil",
+  "262": "Q",
+  "263": "<fi-FI>",
+  "264": "ssa",
+  "265": "lla",
+  "266": "ty",
+  "267": "ki",
+  "268": "pa",
+  "269": "lle",
+  "270": "lu",
+  "271": "tta",
+  "272": "isi",
+  "273": "ise",
+  "274": "tte",
+  "275": "ista",
+  "276": "inen",
+  "277": "llis",
+  "278": "vu",
+  "279": "▁voi",
+  "280": "iden",
+  "281": "ë",
+  "282": "<fr-FR>",
+  "283": "▁est",
+  "284": "▁c",
+  "285": "▁que",
+  "286": "▁à",
+  "287": "es",
+  "288": "▁un",
+  "289": "▁pas",
+  "290": "▁les",
+  "291": "▁qui",
+  "292": "▁il",
+  "293": "▁vous",
+  "294": "▁des",
+  "295": "▁qu",
+  "296": "▁pour",
+  "297": "▁par",
+  "298": "▁ça",
+  "299": "▁une",
+  "300": "ant",
+  "301": "ais",
+  "302": "ez",
+  "303": "▁dans",
+  "304": "▁C",
+  "305": "tre",
+  "306": "ir",
+  "307": "elle",
+  "308": "eur",
+  "309": "▁sur",
+  "310": "▁con",
+  "311": "▁Et",
+  "312": "ement",
+  "313": "tion",
+  "314": "té",
+  "315": "▁tout",
+  "316": "mp",
+  "317": "ique",
+  "318": "▁plus",
+  "319": "eux",
+  "320": "▁dé",
+  "321": "▁fait",
+  "322": "qu",
+  "323": "▁comme",
+  "324": "ac",
+  "325": "ait",
+  "326": "che",
+  "327": "▁mais",
+  "328": "que",
+  "329": "ul",
+  "330": "▁avec",
+  "331": "▁bien",
+  "332": "age",
+  "333": "▁mon",
+  "334": "end",
+  "335": "▁faire",
+  "336": "▁être",
+  "337": "ver",
+  "338": "▁peu",
+  "339": "▁même",
+  "340": "tra",
+  "341": "cha",
+  "342": "▁nous",
+  "343": "▁donc",
+  "344": "▁sont",
+  "345": "▁moi",
+  "346": "ille",
+  "347": "ff",
+  "348": "▁ex",
+  "349": "▁Il",
+  "350": "▁très",
+  "351": "▁cette",
+  "352": "im",
+  "353": "ité",
+  "354": "▁dire",
+  "355": "▁peut",
+  "356": "ance",
+  "357": "aire",
+  "358": "mé",
+  "359": "▁app",
+  "360": "▁aussi",
+  "361": "▁petit",
+  "362": "aux",
+  "363": "▁parce",
+  "364": "onne",
+  "365": "mb",
+  "366": "man",
+  "367": "▁On",
+  "368": "▁quand",
+  "369": "▁autre",
+  "370": "ô",
+  "371": "▁chose",
+  "372": "▁puis",
+  "373": "▁était",
+  "374": "ndre",
+  "375": "port",
+  "376": "▁vraiment",
+  "377": "ence",
+  "378": "▁Mais",
+  "379": "î",
+  "380": "▁avoir",
+  "381": "form",
+  "382": "▁faut",
+  "383": "ign",
+  "384": "▁où",
+  "385": "près",
+  "386": "▁beaucoup",
+  "387": "ture",
+  "388": "û",
+  "389": "â",
+  "390": "<hu-HU>",
+  "391": "▁az",
+  "392": "ok",
+  "393": "gy",
+  "394": "ek",
+  "395": "és",
+  "396": "em",
+  "397": "▁is",
+  "398": "os",
+  "399": "ban",
+  "400": "ik",
+  "401": "cs",
+  "402": "él",
+  "403": "ér",
+  "404": "nek",
+  "405": "ség",
+  "406": "én",
+  "407": "hat",
+  "408": "▁ha",
+  "409": "ép",
+  "410": "vé",
+  "411": "leg",
+  "412": "▁ami",
+  "413": "É",
+  "414": "rend",
+  "415": "<hr-HR>",
+  "416": "▁bi",
+  "417": "▁sa",
+  "418": "ru",
+  "419": "go",
+  "420": "sti",
+  "421": "▁pri",
+  "422": "ima",
+  "423": "nu",
+  "424": "▁pre",
+  "425": "nja",
+  "426": "zi",
+  "427": "vr",
+  "428": "ca",
+  "429": "ba",
+  "430": "<it-IT>",
+  "431": "▁di",
+  "432": "▁che",
+  "433": "co",
+  "434": "▁per",
+  "435": "▁non",
+  "436": "do",
+  "437": "gli",
+  "438": "so",
+  "439": "amo",
+  "440": "sa",
+  "441": "ndo",
+  "442": "fi",
+  "443": "pi",
+  "444": "nti",
+  "445": "tto",
+  "446": "tro",
+  "447": "▁fa",
+  "448": "chi",
+  "449": "▁qua",
+  "450": "bi",
+  "451": "▁del",
+  "452": "pe",
+  "453": "ssi",
+  "454": "▁sono",
+  "455": "nte",
+  "456": "sso",
+  "457": "nta",
+  "458": "cu",
+  "459": "ggi",
+  "460": "sto",
+  "461": "gra",
+  "462": "spe",
+  "463": "cco",
+  "464": "nde",
+  "465": "fe",
+  "466": "pu",
+  "467": "zza",
+  "468": "sci",
+  "469": "▁poi",
+  "470": "sco",
+  "471": "stra",
+  "472": "▁quel",
+  "473": "qui",
+  "474": "sse",
+  "475": "▁inter",
+  "476": "sce",
+  "477": "▁com",
+  "478": "▁comp",
+  "479": "▁prima",
+  "480": "▁imp",
+  "481": "▁cui",
+  "482": "scri",
+  "483": "<lt-LT>",
+  "484": "ai",
+  "485": "▁ir",
+  "486": "as",
+  "487": "▁tai",
+  "488": "uo",
+  "489": "tin",
+  "490": "▁vis",
+  "491": "ly",
+  "492": "gal",
+  "493": "jo",
+  "494": "tar",
+  "495": "yb",
+  "496": "▁Ir",
+  "497": "▁Tai",
+  "498": "▁nu",
+  "499": "▁mes",
+  "500": "▁pra",
+  "501": "<lv-LV>",
+  "502": "▁pie",
+  "503": "ies",
+  "504": "ot",
+  "505": "▁Un",
+  "506": "▁Ta",
+  "507": "▁lie",
+  "508": "dar",
+  "509": "sim",
+  "510": "dien",
+  "511": "gan",
+  "512": "▁ap",
+  "513": "▁nav",
+  "514": "▁cit",
+  "515": "▁Ne",
+  "516": "zin",
+  "517": "reiz",
+  "518": "▁20",
+  "519": "<nl-NL>",
+  "520": "▁dat",
+  "521": "▁we",
+  "522": "▁En",
+  "523": "▁dan",
+  "524": "▁met",
+  "525": "▁wat",
+  "526": "der",
+  "527": "ui",
+  "528": "den",
+  "529": "op",
+  "530": "oor",
+  "531": "▁of",
+  "532": "ven",
+  "533": "acht",
+  "534": "ond",
+  "535": "▁dit",
+  "536": "▁wil",
+  "537": "▁We",
+  "538": "vol",
+  "539": "llen",
+  "540": "▁Dan",
+  "541": "▁Nou",
+  "542": "<pl-PL>",
+  "543": "nie",
+  "544": "wa",
+  "545": "my",
+  "546": "cie",
+  "547": "nia",
+  "548": "wo",
+  "549": "rze",
+  "550": "by",
+  "551": "za",
+  "552": "dy",
+  "553": "ry",
+  "554": "mie",
+  "555": "pie",
+  "556": "uje",
+  "557": "▁bo",
+  "558": "bie",
+  "559": "sy",
+  "560": "▁Po",
+  "561": "ski",
+  "562": "▁No",
+  "563": "▁Na",
+  "564": "▁musi",
+  "565": "<pt-BR>",
+  "566": "pt",
+  "567": "-",
+  "568": "▁eu",
+  "569": "▁para",
+  "570": "▁pe",
+  "571": "▁tem",
+  "572": "▁O",
+  "573": "pre",
+  "574": "ria",
+  "575": "▁fo",
+  "576": "mos",
+  "577": "bo",
+  "578": "▁por",
+  "579": "▁Eu",
+  "580": "be",
+  "581": "▁esse",
+  "582": "ente",
+  "583": "▁mas",
+  "584": "qua",
+  "585": "fica",
+  "586": "▁vou",
+  "587": "▁Se",
+  "588": "▁fala",
+  "589": "...",
+  "590": "ço",
+  "591": "▁Por",
+  "592": "▁Co",
+  "593": "iza",
+  "594": "<ro-RO>",
+  "595": "▁mai",
+  "596": "sc",
+  "597": "are",
+  "598": "▁din",
+  "599": "rea",
+  "600": "du",
+  "601": "▁M",
+  "602": "▁fac",
+  "603": "lor",
+  "604": "▁mult",
+  "605": "per",
+  "606": "cur",
+  "607": "tor",
+  "608": "inte",
+  "609": "▁sau",
+  "610": "tat",
+  "611": "ori",
+  "612": "▁prim",
+  "613": "▁lui",
+  "614": "▁sub",
+  "615": "▁prin",
+  "616": "▁alt",
+  "617": "stru",
+  "618": "fer",
+  "619": "<ru-RU>",
+  "620": "<sk-SK>",
+  "621": "sk",
+  "622": "▁aj",
+  "623": "ob",
+  "624": "ali",
+  "625": "nov",
+  "626": "rov",
+  "627": "rob",
+  "628": "▁spo",
+  "629": "<sl-SL>",
+  "630": "sl",
+  "631": "udi",
+  "632": "del",
+  "633": "▁sem",
+  "634": "▁samo",
+  "635": "▁let",
+  "636": "▁Pre",
+  "637": "▁prot",
+  "638": "▁internet",
+  "639": "▁film",
+  "640": "<sv-SE>",
+  "641": "▁att",
+  "642": "▁inte",
+  "643": "▁av",
+  "644": "all",
+  "645": "era",
+  "646": "pp",
+  "647": "sam",
+  "648": "▁vill",
+  "649": "erna",
+  "650": "ande",
+  "651": "ade",
+  "652": "▁hur",
+  "653": "bil",
+  "654": "▁min",
+  "655": "lev",
+  "656": "land",
+  "657": "text",
+  "658": "person",
+  "659": "<uk-UA>",
+  "660": "<ar-AR>",
+  "661": "'",
+  "662": "▁H",
+  "663": "▁D",
+  "664": "aus",
+  "665": "▁N",
+  "666": "▁Be",
+  "667": "mm",
+  "668": "ab",
+  "669": "▁Er",
+  "670": "ssen",
+  "671": "hl",
+  "672": "rie",
+  "673": "▁An",
+  "674": "rau",
+  "675": "▁So",
+  "676": "▁Aus",
+  "677": "<en-GB>",
+  "678": "▁can",
+  "679": "ed",
+  "680": "ay",
+  "681": "th",
+  "682": "ic",
+  "683": "hi",
+  "684": "▁not",
+  "685": "ight",
+  "686": "ex",
+  "687": "▁four",
+  "688": "▁great",
+  "689": "▁six",
+  "690": "ill",
+  "691": "▁but",
+  "692": "▁got",
+  "693": "▁don",
+  "694": "▁out",
+  "695": "▁from",
+  "696": "▁fine",
+  "697": "able",
+  "698": "ever",
+  "699": "▁month",
+  "700": "▁question",
+  "701": "▁possible",
+  "702": "▁moment",
+  "703": "▁change",
+  "704": "<en-US>",
+  "705": "ow",
+  "706": "ive",
+  "707": "ate",
+  "708": "ad",
+  "709": "ect",
+  "710": "ng",
+  "711": "ther",
+  "712": "act",
+  "713": "ist",
+  "714": "▁part",
+  "715": "side",
+  "716": "\"",
+  "717": "<es-ES>",
+  "718": "ción",
+  "719": "res",
+  "720": "▁La",
+  "721": "dos",
+  "722": "▁las",
+  "723": "men",
+  "724": "par",
+  "725": "rio",
+  "726": "cho",
+  "727": "ble",
+  "728": "▁Ca",
+  "729": "▁Su",
+  "730": "▁son",
+  "731": "▁Con",
+  "732": "ones",
+  "733": "ecto",
+  "734": "▁Com",
+  "735": "<es-US>",
+  "736": "cia",
+  "737": "▁Y",
+  "738": "ron",
+  "739": "les",
+  "740": "bu",
+  "741": "<fr-CA>",
+  "742": "ré",
+  "743": "ée",
+  "744": "▁Les",
+  "745": "our",
+  "746": "▁Ce",
+  "747": "com",
+  "748": "▁Elle",
+  "749": "▁Cet",
+  "750": "ux",
+  "751": "ale",
+  "752": "ction",
+  "753": "▁pré",
+  "754": "▁deux",
+  "755": "if",
+  "756": "lé",
+  "757": "ère",
+  "758": "ière",
+  "759": "iste",
+  "760": "▁parti",
+  "761": "▁été",
+  "762": "avec",
+  "763": "jour",
+  "764": "app",
+  "765": "cul",
+  "766": "▁égale",
+  "767": "aine",
+  "768": "gue",
+  "769": "▁nombre",
+  "770": "▁étai",
+  "771": "tout",
+  "772": "▁grand",
+  "773": "▁commun",
+  "774": "Une",
+  "775": "œ",
+  "776": "ï",
+  "777": "À",
+  "778": "Â",
+  "779": "av",
+  "780": "ouv",
+  "781": "êt",
+  "782": "ois",
+  "783": "pri",
+  "784": "voir",
+  "785": "sion",
+  "786": "ix",
+  "787": "était",
+  "788": "ard",
+  "789": "aient",
+  "790": "<he-IL>",
+  "791": "<hi-IN>",
+  "792": "!",
+  "793": "<ja-JP>",
+  "794": "<ko-KR>",
+  "795": "<nb-NO>",
+  "796": "▁ble",
+  "797": "opp",
+  "798": "▁Han",
+  "799": "▁han",
+  "800": "<nn-NO>",
+  "801": "eg",
+  "802": "kk",
+  "803": "dig",
+  "804": "ord",
+  "805": "▁tru",
+  "806": "▁sei",
+  "807": "ller",
+  "808": "▁fle",
+  "809": "car",
+  "810": "▁foi",
+  "811": "ito",
+  "812": "ram",
+  "813": "fa",
+  "814": "▁meu",
+  "815": "nça",
+  "816": "▁mil",
+  "817": "▁passa",
+  "818": "▁casa",
+  "819": "<pt-PT>",
+  "820": "das",
+  "821": "▁Pa",
+  "822": "forma",
+  "823": "mar",
+  "824": "▁entre",
+  "825": "fun",
+  "826": "▁grande",
+  "827": "▁Este",
+  "828": "var",
+  "829": "<th-TH>",
+  "830": "<tr-TR>",
+  "831": "ş",
+  "832": "ğ",
+  "833": "ya",
+  "834": "▁ve",
+  "835": "lar",
+  "836": "ler",
+  "837": "ye",
+  "838": "▁bu",
+  "839": "lan",
+  "840": "ara",
+  "841": "▁Bu",
+  "842": "inde",
+  "843": "yo",
+  "844": "<zh-CN>",
+  "845": "<vi-VN>",
+  "846": "▁t",
+  "847": "nh",
+  "848": "<blank>"
+}
\ No newline at end of file