Upload 352 files

it/1120ms/decoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4cdca6bf678463f31354072f526088e5bdf5115ae94c04e387bb35b2c7a607d6
+size 243

it/1120ms/decoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c993f8b96ce22027cd2ed42d99b7e61f93a01197bb17cadada8eb989e946dec
+size 433

it/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, [806, 640]> module_prediction_embed_weight_to_fp16 = const()[name = string("module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [806, 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(1031808)))];
+            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(4308672)))];
+            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(7585536)))];
+            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(7590720)))];
+            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(10867584)))];
+            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(14144448)))];
+            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);
+}

it/1120ms/decoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1360d93c68c3e9c54bda4adaec860753949f3b0dc93bc98f4edc9d6f8dd5595c
+size 14149632

it/1120ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:73bb3afa62698bc822b6d32b3731d0bc40521e03737e3139e10a768542fca1fe
+size 10359

it/1120ms/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1360d93c68c3e9c54bda4adaec860753949f3b0dc93bc98f4edc9d6f8dd5595c
+size 14149632

it/1120ms/decoder.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "7CBCED8D-FA6A-45B0-BF60-30DB0A653074": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "AFD197FC-BECC-451A-961C-C0CA05D58065": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "AFD197FC-BECC-451A-961C-C0CA05D58065"
+}

it/1120ms/decoder_joint.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:12f4bcf5114baa2b3a37b8ebeab6c519109bd857e50ec345c458b7a6c4deb20e
+size 243

it/1120ms/decoder_joint.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53754e2eaa7e0f7435220b47b621c5f3d8c5f2da83edd46efa5950fa723ef1d9
+size 454

it/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, [806, 640]> decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [806, 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(1031808)))];
+            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(4308672)))];
+            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(7585536)))];
+            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(7590720)))];
+            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(10867584)))];
+            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(14144448)))];
+            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(14149632)))];
+            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(15460416)))];
+            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(15461760)))];
+            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(16281024)))];
+            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, [806, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [806, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16282368)))];
+            tensor<fp16, [806]> joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [806]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17314112)))];
+            tensor<fp16, [1, 1, 1, 806]> 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, 806]> 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);
+}

it/1120ms/decoder_joint.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e507a69196a04e30adfafc302b6a5f5f527e45c1965c65dd81d63a621cae2064
+size 17315788

it/1120ms/decoder_joint.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5fa6a6d89c1c07ae16f162f5a3b6809b12aafe57a663f5cdb270be3dec7b1427
+size 13745

it/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:e507a69196a04e30adfafc302b6a5f5f527e45c1965c65dd81d63a621cae2064
+size 17315788

it/1120ms/decoder_joint.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "2B19A50C-1D16-4D97-BE3C-D9BCF35884CF": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "9CA734BC-CFD2-4F39-B068-BE69ABCAAD1F": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "2B19A50C-1D16-4D97-BE3C-D9BCF35884CF"
+}

it/1120ms/decoder_joint_noencproj.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40ec657603479e7dbf8cdb3d6368349eb8b766a52439a26a735d1fadf1b4281d
+size 243

it/1120ms/decoder_joint_noencproj.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c682ba0e028fb7ab6557f8ac1006febc8ec8dd81e4ef8d3a2c05d876e2dbcc8e
+size 519

it/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, [806, 640]> decoder_module_prediction_embed_weight_to_fp16 = const()[name = string("decoder_module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [806, 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(806)];
+            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(1031808)))];
+            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(4308672)))];
+            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(7585536)))];
+            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(7590720)))];
+            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(10867584)))];
+            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(14144448)))];
+            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(14149632)))];
+            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(14968896)))];
+            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, [806, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [806, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14970240)))];
+            tensor<fp16, [806]> joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [806]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16001984)))];
+            tensor<fp16, [1, 1, 1, 806]> 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, 806]> 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, 806]> 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, 806]> 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);
+}

it/1120ms/decoder_joint_noencproj.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ca6a467ebf44612a8032c6c4ddf323e35a9ffaa15c01822528fb97144ec9439
+size 16003660

it/1120ms/decoder_joint_noencproj.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a46ef6436e8c1e5af6f379658d871f2a40dff3a13c946d4f8df736b19cdc45e
+size 14630

it/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 sha256:5ca6a467ebf44612a8032c6c4ddf323e35a9ffaa15c01822528fb97144ec9439
+size 16003660

it/1120ms/decoder_joint_noencproj.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "E92AE661-EE59-447C-B9BE-52D50BC4FC37": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "F7138F3F-8633-4C54-A3F4-D0F6E8220E39": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "F7138F3F-8633-4C54-A3F4-D0F6E8220E39"
+}

it/1120ms/encoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0638780288ed25026170632a153dbed5d798954deac843fa58ddaae3190914d4
+size 243

it/1120ms/encoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f76d62779c43fa5272f8977cfc6d3e368abd5cc38841834d05834021633dd2f
+size 662

it/1120ms/encoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:011a5342cc1e2901633a8f0468be561f50dfc289ebe51851cddb4831f9a6a23f
+size 565952640

it/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

it/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

it/1120ms/encoder.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "2A6DB86E-A930-4F02-B70E-CA03D7F3AD84": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "712F619E-4FA1-4165-A290-6C1344B0A816": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "2A6DB86E-A930-4F02-B70E-CA03D7F3AD84"
+}

it/1120ms/joint.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cae47ef3b8cb372625305f4cdd31159b5ce56be1470c94613ff946be342f6d7a
+size 243

it/1120ms/joint.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:215d8dd2e33da0c37f08e6b0c0a1e997a3e056f2cb9113fdcf17f8027a61216d
+size 341

it/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, [806, 640]> module_joint_net_2_weight_to_fp16 = const()[name = string("module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [806, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2132800)))];
+            tensor<fp16, [806]> module_joint_net_2_bias_to_fp16 = const()[name = string("module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [806]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3164544)))];
+            tensor<fp16, [1, 1, 1, 806]> 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, 806]> logits = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = string("cast_0")];
+        } -> (logits);
+}

it/1120ms/joint.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1740c99cbe34ebeaa0163c8421135b4586df09960ef07fe02abb2a94b5693411
+size 3166220

it/1120ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d902ae83b8a93f0f4240a4a6939466dbd1a6b2291f1615d81d7ac26d9115bc23
+size 4486

it/1120ms/joint.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1740c99cbe34ebeaa0163c8421135b4586df09960ef07fe02abb2a94b5693411
+size 3166220

it/1120ms/joint.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "6D4EAD3B-A17D-4807-8572-3E0A513C0C7E": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "8242AF76-3606-4A98-8B3D-E0BB730201A4": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "6D4EAD3B-A17D-4807-8572-3E0A513C0C7E"
+}

it/1120ms/joint_noencproj_batched.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24510a0ca90eea355a0e446ba54d63004a3f2c51d30eb3e415e4af938ac1acba
+size 243

it/1120ms/joint_noencproj_batched.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc93bb8f241754f1359837f597d600b029d687ce0c57a4280fa586f8c8386337
+size 406

it/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, [806, 640]> joint_module_joint_net_2_weight_to_fp16 = const()[name = string("joint_module_joint_net_2_weight_to_fp16"), val = tensor<fp16, [806, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(820672)))];
+            tensor<fp16, [806]> joint_module_joint_net_2_bias_to_fp16 = const()[name = string("joint_module_joint_net_2_bias_to_fp16"), val = tensor<fp16, [806]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1852416)))];
+            tensor<fp16, [1, 4, 1, 806]> 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, 806]> logits = cast(dtype = linear_1_cast_fp16_to_fp32_dtype_0, x = linear_1_cast_fp16)[name = string("cast_0")];
+        } -> (logits);
+}

it/1120ms/joint_noencproj_batched.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd83e82dcfec315f28c8a8872b0d7f22e668a2c485821de86a0379ae2b3864ad
+size 1854092

it/1120ms/joint_noencproj_batched.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26a19f2e269811070e57eda978e8226ca9dcca6177b46b95f916538b5f815716
+size 3839

it/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:cd83e82dcfec315f28c8a8872b0d7f22e668a2c485821de86a0379ae2b3864ad
+size 1854092

it/1120ms/joint_noencproj_batched.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "4233CE8E-FB95-4FF9-BCD8-2A834D55C580": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "96E0F26C-90DC-49EE-B510-D0FB3FC812CC": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "96E0F26C-90DC-49EE-B510-D0FB3FC812CC"
+}

it/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": 805,
+  "blank_idx": 805,
+  "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,
+    63,
+    115,
+    167,
+    226,
+    227,
+    259,
+    276,
+    328,
+    353,
+    368,
+    462,
+    481,
+    499,
+    518,
+    542,
+    571,
+    602,
+    603,
+    612,
+    624,
+    646,
+    647,
+    667,
+    689,
+    699,
+    720,
+    727,
+    747,
+    748,
+    750,
+    751,
+    752,
+    756,
+    774,
+    787,
+    788,
+    801,
+    802
+  ]
+}

it/1120ms/preprocessor.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:97be07beee7a6a28b7db85d47087d5b018ebcd1fc0b1565707141d574244bdc9
+size 243

it/1120ms/preprocessor.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7b11e08aba46d1845d8ad3f247717e0f6fae35b21d71d52e44a69ea73587bfe
+size 371

it/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);
+}

it/1120ms/preprocessor.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297514e2b211d14b0e53cb97193d679bb89ead98d28e578f3f1d049ddbcc36b3
+size 592384

it/1120ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f1df1cad6f43ce4fa090e0f9a33bb5ddf25a0aaeca2be136339878b5b9de45c0
+size 15878

it/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

it/1120ms/preprocessor.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "162F4F34-AB45-4D09-8F9D-2F76AAE79E0F": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "B5F24360-A5E1-4F41-9872-3D77094622F0": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "B5F24360-A5E1-4F41-9872-3D77094622F0"
+}