15s, moved Transpose. Not quantized

Decoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4238c4e81ecd0dc94bd7dfbb60f7e2cc824107c1ffe0387b8607b72833dba350
+size 243

Decoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18647af085d87bd8f3121c8a9b4d4564c1ede038dab63d295b4e745cf2d7fb99
+size 554

Decoder.mlmodelc/metadata.json

@@ -0,0 +1,123 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Parakeet decoder (RNNT prediction network)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 640 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 640, 1]",
+        "name" : "decoder",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "h_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "c_out",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Float16",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Select" : 1,
+      "Ios17.squeeze" : 4,
+      "Ios17.gather" : 1,
+      "Ios17.cast" : 8,
+      "Ios17.lstm" : 2,
+      "Split" : 2,
+      "Ios17.add" : 1,
+      "Ios17.transpose" : 2,
+      "Ios17.greaterEqual" : 1,
+      "Identity" : 1,
+      "Stack" : 2
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int16, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1]",
+        "name" : "targets",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "target_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "h_in",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "c_in",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2025-09-19",
+      "com.github.apple.coremltools.source" : "torch==2.7.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "parakeet_decoder",
+    "method" : "predict"
+  }
+]

Decoder.mlmodelc/model.mil

@@ -0,0 +1,73 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.7.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0b1"}})]
+{
+    func main<ios17>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1]> target_length, tensor<int32, [1, 1]> targets) {
+            tensor<int32, []> y_batch_dims_0 = const()[name = tensor<string, []>("y_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> y_validate_indices_0 = const()[name = tensor<string, []>("y_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [8193, 640]> module_prediction_embed_weight_to_fp16 = const()[name = tensor<string, []>("module_prediction_embed_weight_to_fp16"), val = tensor<fp16, [8193, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<string, []> targets_to_int16_dtype_0 = const()[name = tensor<string, []>("targets_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<string, []> cast_1_dtype_0 = const()[name = tensor<string, []>("cast_1_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, []> greater_equal_0_y_0 = const()[name = tensor<string, []>("greater_equal_0_y_0"), val = tensor<int32, []>(0)];
+            tensor<int16, [1, 1]> targets_to_int16 = cast(dtype = targets_to_int16_dtype_0, x = targets)[name = tensor<string, []>("cast_9")];
+            tensor<int32, [1, 1]> cast_1 = cast(dtype = cast_1_dtype_0, x = targets_to_int16)[name = tensor<string, []>("cast_8")];
+            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = cast_1, y = greater_equal_0_y_0)[name = tensor<string, []>("greater_equal_0")];
+            tensor<int32, []> slice_by_index_0 = const()[name = tensor<string, []>("slice_by_index_0"), val = tensor<int32, []>(8193)];
+            tensor<int32, [1, 1]> add_2 = add(x = cast_1, y = slice_by_index_0)[name = tensor<string, []>("add_2")];
+            tensor<int32, [1, 1]> select_0 = select(a = cast_1, b = add_2, cond = greater_equal_0)[name = tensor<string, []>("select_0")];
+            tensor<int32, []> y_cast_fp16_cast_uint16_axis_0 = const()[name = tensor<string, []>("y_cast_fp16_cast_uint16_axis_0"), val = tensor<int32, []>(0)];
+            tensor<string, []> select_0_to_int16_dtype_0 = const()[name = tensor<string, []>("select_0_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 1]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = tensor<string, []>("cast_7")];
+            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 = module_prediction_embed_weight_to_fp16)[name = tensor<string, []>("y_cast_fp16_cast_uint16_cast_uint16")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, []> split_0_num_splits_0 = const()[name = tensor<string, []>("split_0_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> split_0_axis_0 = const()[name = tensor<string, []>("split_0_axis_0"), val = tensor<int32, []>(0)];
+            tensor<string, []> h_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("h_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [2, 1, 640]> h_in_to_fp16 = cast(dtype = h_in_to_fp16_dtype_0, x = h_in)[name = tensor<string, []>("cast_6")];
+            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 = tensor<string, []>("split_0_cast_fp16")];
+            tensor<int32, []> split_1_num_splits_0 = const()[name = tensor<string, []>("split_1_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> split_1_axis_0 = const()[name = tensor<string, []>("split_1_axis_0"), val = tensor<int32, []>(0)];
+            tensor<string, []> c_in_to_fp16_dtype_0 = const()[name = tensor<string, []>("c_in_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [2, 1, 640]> c_in_to_fp16 = cast(dtype = c_in_to_fp16_dtype_0, x = c_in)[name = tensor<string, []>("cast_5")];
+            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 = tensor<string, []>("split_1_cast_fp16")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = tensor<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 = tensor<string, []>("input_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = tensor<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 = tensor<string, []>("input_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> input_lstm_layer_0_direction_0 = const()[name = tensor<string, []>("input_lstm_layer_0_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_lstm_layer_0_output_sequence_0 = const()[name = tensor<string, []>("input_lstm_layer_0_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_lstm_layer_0_recurrent_activation_0 = const()[name = tensor<string, []>("input_lstm_layer_0_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_lstm_layer_0_cell_activation_0 = const()[name = tensor<string, []>("input_lstm_layer_0_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_lstm_layer_0_activation_0 = const()[name = tensor<string, []>("input_lstm_layer_0_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = tensor<string, []>("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10487168)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = tensor<string, []>("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13764032)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = tensor<string, []>("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17040896)))];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = transpose(perm = input_3_perm_0, x = y_cast_fp16_cast_uint16_cast_uint16)[name = tensor<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 = tensor<string, []>("input_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> input_lstm_h0_squeeze_axes_0 = const()[name = tensor<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 = tensor<string, []>("input_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> input_lstm_c0_squeeze_axes_0 = const()[name = tensor<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 = tensor<string, []>("input_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> input_direction_0 = const()[name = tensor<string, []>("input_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_output_sequence_0 = const()[name = tensor<string, []>("input_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_recurrent_activation_0 = const()[name = tensor<string, []>("input_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_cell_activation_0 = const()[name = tensor<string, []>("input_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_activation_0 = const()[name = tensor<string, []>("input_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = tensor<string, []>("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17046080)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = tensor<string, []>("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20322944)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = tensor<string, []>("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23599808)))];
+            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 = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, []> obj_3_axis_0 = const()[name = tensor<string, []>("obj_3_axis_0"), val = tensor<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 = tensor<string, []>("obj_3_cast_fp16")];
+            tensor<string, []> obj_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_3_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, []> obj_axis_0 = const()[name = tensor<string, []>("obj_axis_0"), val = tensor<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 = tensor<string, []>("obj_cast_fp16")];
+            tensor<string, []> obj_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("obj_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = tensor<string, []>("transpose_0_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
+            tensor<string, []> transpose_0_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("transpose_0_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp16, [1, 640, 1]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = input_cast_fp16_0)[name = tensor<string, []>("transpose_1")];
+            tensor<fp32, [1, 640, 1]> decoder = cast(dtype = transpose_0_cast_fp16_to_fp32_dtype_0, x = transpose_0_cast_fp16)[name = tensor<string, []>("cast_2")];
+            tensor<fp32, [2, 1, 640]> c_out = cast(dtype = obj_cast_fp16_to_fp32_dtype_0, x = obj_cast_fp16)[name = tensor<string, []>("cast_3")];
+            tensor<fp32, [2, 1, 640]> h_out = cast(dtype = obj_3_cast_fp16_to_fp32_dtype_0, x = obj_3_cast_fp16)[name = tensor<string, []>("cast_4")];
+            tensor<int32, [1]> target_length_tmp = identity(x = target_length)[name = tensor<string, []>("target_length_tmp")];
+        } -> (decoder, h_out, c_out);
+}

Decoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48adf0f0d47c406c8253d4f7fef967436a39da14f5a65e66d5a4b407be355d41
+size 23604992