Upload folder using huggingface_hub

config.json

@@ -0,0 +1,92 @@
+{
+  "model_type": "fastconformer_rnnt",
+  "source_model": "nvidia/nemotron-speech-streaming-en-0.6b",
+  "format": "coreml",
+
+  "audio": {
+    "sample_rate": 16000,
+    "sample_format": "S16_LE",
+    "bytes_per_second": 32000
+  },
+
+  "preprocessor": {
+    "n_mels": 128,
+    "n_fft": 512,
+    "hop_length": 160,
+    "win_length": 400,
+    "window": "hann",
+    "preemph": 0.97,
+    "dither": 1e-05,
+    "normalize": null,
+    "pad_to": 0,
+    "mel_norm": "slaney",
+    "mel_layout": "band_major",
+    "_comment": "mel_layout=band_major means output shape is [n_mels, n_frames], not [n_frames, n_mels]. Mel preprocessing runs on the host (not in CoreML)."
+  },
+
+  "encoder": {
+    "model_file": "encoder.mlmodelc",
+    "layers": 24,
+    "dim": 1024,
+    "chunk_mel_frames": 56,
+    "pre_encode_cache_frames": 9,
+    "total_input_frames": 65,
+    "compute_units": "CPU_AND_NE",
+    "compute_precision": "FLOAT16",
+    "cache_last_channel_shape": [1, 24, 70, 1024],
+    "cache_last_time_shape": [1, 24, 1024, 8],
+    "_comment_cache_layout": "batch-first [B, L, ...] for CoreML. NeMo uses layer-first [L, B, ...] internally; the wrapper transposes.",
+    "inputs": {
+      "audio_signal": {"dtype": "float32", "shape": [1, 128, 65]},
+      "cache_last_channel": {"dtype": "float32", "shape": [1, 24, 70, 1024], "init": "zeros"},
+      "cache_last_time": {"dtype": "float32", "shape": [1, 24, 1024, 8], "init": "zeros"},
+      "cache_last_channel_len": {"dtype": "int32", "shape": [1], "init": "zeros"}
+    },
+    "outputs": {
+      "encoded": {"dtype": "float16", "shape": [1, 1024, 7]},
+      "encoded_length": {"dtype": "int32", "shape": [1]},
+      "cache_channel_out": {"dtype": "float16", "shape": [1, 24, 70, 1024], "_comment": "feed back as cache_last_channel"},
+      "cache_time_out": {"dtype": "float16", "shape": [1, 24, 1024, 8], "_comment": "feed back as cache_last_time"},
+      "cache_len_out": {"dtype": "int32", "shape": [1], "_comment": "feed back as cache_last_channel_len"}
+    },
+    "_comment_strides": "CoreML output MLMultiArrays may have non-contiguous strides due to ANE alignment padding (e.g. stride=32 for dim size 7). Callers must use stride-aware copy, not flat memcpy."
+  },
+
+  "decoder": {
+    "model_file": "decoder.mlmodelc",
+    "prediction_layers": 2,
+    "prediction_hidden": 640,
+    "vocab_size": 1025,
+    "blank_id": 1024,
+    "max_symbols_per_frame": 10,
+    "compute_units": "CPU_ONLY",
+    "compute_precision": "FLOAT16",
+    "inputs": {
+      "encoder_outputs": {"dtype": "float32", "shape": [1, 1024, 1], "_comment": "single encoder frame"},
+      "targets": {"dtype": "int32", "shape": [1, 1], "_comment": "previous non-blank token (blank_id for first)"},
+      "target_length": {"dtype": "int32", "shape": [1], "value": 1},
+      "input_states_1": {"dtype": "float32", "shape": [2, 1, 640], "init": "zeros"},
+      "input_states_2": {"dtype": "float32", "shape": [2, 1, 640], "init": "zeros"}
+    },
+    "outputs": {
+      "outputs": {"dtype": "float16", "shape": [1, 1025], "_comment": "logits over vocab, argmax to get token"},
+      "output_states_1": {"dtype": "float16", "shape": [2, 1, 640], "_comment": "feed back as input_states_1"},
+      "output_states_2": {"dtype": "float16", "shape": [2, 1, 640], "_comment": "feed back as input_states_2"}
+    }
+  },
+
+  "streaming": {
+    "chunk_duration_ms": 560,
+    "chunk_audio_samples": 8960,
+    "algorithm": "Cache-aware streaming FastConformer. Feed 560ms audio chunks. For each chunk: compute mel (56 frames), prepend 9 cached frames, run encoder. Then for each encoder output frame, run RNNT greedy decode: feed one encoder frame to decoder, argmax logits, if not blank emit token and loop (up to max_symbols_per_frame), if blank move to next frame. Feed cache outputs back as next chunk's cache inputs. Feed decoder states back for next symbol/frame."
+  },
+
+  "ane_profile": {
+    "encoder_ane_ops": 1486,
+    "encoder_cpu_ops": 111,
+    "encoder_ane_pct": 93.0,
+    "cpu_op_breakdown": "select (72, attention masking), softmax (24, one per conformer layer), expand_dims/logical/cast (15, control flow)",
+    "decoder": "100% CPU (CPU_ONLY compute units)",
+    "inference_speed": "~15.5x realtime on Apple Silicon (M-series)"
+  }
+}

fp16/decoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:89c4f741f02efc6cb362c7dc94893fb832783cc608d54542752a74653c36cb2f
+size 243

fp16/decoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0711cd32735530249dba2d75b3b9f57fd5d3d25b215bcef244c2e949f194a41f
+size 565

fp16/decoder.mlmodelc/metadata.json

@@ -0,0 +1,134 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 1025)",
+        "shortDescription" : "",
+        "shape" : "[1, 1025]",
+        "name" : "outputs",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "output_states_1",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "output_states_2",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.linear" : 3,
+      "Ios18.expandDims" : 2,
+      "Select" : 1,
+      "Ios18.relu" : 1,
+      "Ios18.lstm" : 2,
+      "Ios18.gather" : 1,
+      "Stack" : 2,
+      "Ios18.add" : 2,
+      "Ios18.transpose" : 3,
+      "Ios18.cast" : 6,
+      "Ios18.squeeze" : 6,
+      "Ios18.greaterEqual" : 1,
+      "Split" : 2,
+      "Identity" : 1
+    },
+    "computePrecision" : "Mixed (Float16, Int16, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "15.0",
+      "tvOS" : "18.0",
+      "visionOS" : "2.0",
+      "watchOS" : "11.0",
+      "iOS" : "18.0",
+      "macCatalyst" : "18.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-03-22",
+      "com.github.apple.coremltools.source" : "torch==2.7.1",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1024 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1024, 1]",
+        "name" : "encoder_outputs",
+        "type" : "MultiArray"
+      },
+      {
+        "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" : "input_states_1",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "input_states_2",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "decoder",
+    "method" : "predict"
+  }
+]

fp16/decoder.mlmodelc/model.mil

@@ -0,0 +1,90 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.7.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0b1"}})]
+{
+    func main<ios18>(tensor<fp32, [1, 1024, 1]> encoder_outputs, tensor<fp32, [2, 1, 640]> input_states_1, tensor<fp32, [2, 1, 640]> input_states_2, tensor<int32, [1]> target_length, tensor<int32, [1, 1]> targets) {
+            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, [1025, 640]> decoder_prediction_embed_weight_to_fp16 = const()[name = string("decoder_prediction_embed_weight_to_fp16"), val = tensor<fp16, [1025, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string targets_to_int16_dtype_0 = const()[name = string("targets_to_int16_dtype_0"), val = string("int16")];
+            string cast_4_dtype_0 = const()[name = string("cast_4_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]> targets_to_int16 = cast(dtype = targets_to_int16_dtype_0, x = targets)[name = string("cast_10")];
+            tensor<int32, [1, 1]> cast_4 = cast(dtype = cast_4_dtype_0, x = targets_to_int16)[name = string("cast_9")];
+            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = cast_4, 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(1025)];
+            tensor<int32, [1, 1]> add_2 = add(x = cast_4, y = slice_by_index_0)[name = string("add_2")];
+            tensor<int32, [1, 1]> select_0 = select(a = cast_4, 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_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 input_states_1_to_fp16_dtype_0 = const()[name = string("input_states_1_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> input_states_1_to_fp16 = cast(dtype = input_states_1_to_fp16_dtype_0, x = input_states_1)[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 = input_states_1_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 input_states_2_to_fp16_dtype_0 = const()[name = string("input_states_2_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [2, 1, 640]> input_states_2_to_fp16 = cast(dtype = input_states_2_to_fp16_dtype_0, x = input_states_2)[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 = input_states_2_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(1312128)))];
+            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(4588992)))];
+            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(7865856)))];
+            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_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(7871040)))];
+            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(11147904)))];
+            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(14424768)))];
+            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]> output_states_1 = 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")];
+            int32 obj_axis_0 = const()[name = string("obj_axis_0"), val = int32(0)];
+            tensor<fp16, [2, 1, 640]> output_states_2 = stack(axis = obj_axis_0, values = (input_5_lstm_layer_0_cast_fp16_2, input_5_cast_fp16_2))[name = string("obj_cast_fp16")];
+            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_outputs_to_fp16_dtype_0 = const()[name = string("encoder_outputs_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [640, 1024]> joint_enc_weight_to_fp16 = const()[name = string("joint_enc_weight_to_fp16"), val = tensor<fp16, [640, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14429952)))];
+            tensor<fp16, [640]> joint_enc_bias_to_fp16 = const()[name = string("joint_enc_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15740736)))];
+            tensor<fp16, [1, 1024, 1]> encoder_outputs_to_fp16 = cast(dtype = encoder_outputs_to_fp16_dtype_0, x = encoder_outputs)[name = string("cast_5")];
+            tensor<fp16, [1, 1, 1024]> input_7_cast_fp16 = transpose(perm = input_7_perm_0, x = encoder_outputs_to_fp16)[name = string("transpose_2")];
+            tensor<fp16, [1, 1, 640]> linear_0_cast_fp16 = linear(bias = joint_enc_bias_to_fp16, weight = joint_enc_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<fp16, [640, 640]> joint_pred_weight_to_fp16 = const()[name = string("joint_pred_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(15742080)))];
+            tensor<fp16, [640]> joint_pred_bias_to_fp16 = const()[name = string("joint_pred_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16561344)))];
+            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_pred_bias_to_fp16, weight = joint_pred_weight_to_fp16, x = transpose_1_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<int32, [1]> var_78_axes_0 = const()[name = string("op_78_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 640]> var_78_cast_fp16 = expand_dims(axes = var_78_axes_0, x = linear_0_cast_fp16)[name = string("op_78_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_78_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, [1025, 640]> joint_joint_net_2_weight_to_fp16 = const()[name = string("joint_joint_net_2_weight_to_fp16"), val = tensor<fp16, [1025, 640]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16562688)))];
+            tensor<fp16, [1025]> joint_joint_net_2_bias_to_fp16 = const()[name = string("joint_joint_net_2_bias_to_fp16"), val = tensor<fp16, [1025]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17874752)))];
+            tensor<fp16, [1, 1, 1, 1025]> linear_2_cast_fp16 = linear(bias = joint_joint_net_2_bias_to_fp16, weight = joint_joint_net_2_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_2_cast_fp16")];
+            tensor<int32, [1]> var_91_axes_0 = const()[name = string("op_91_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1025]> var_91_cast_fp16 = squeeze(axes = var_91_axes_0, x = linear_2_cast_fp16)[name = string("op_91_cast_fp16")];
+            tensor<int32, [1]> var_93_axes_0 = const()[name = string("op_93_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1025]> outputs = squeeze(axes = var_93_axes_0, x = var_91_cast_fp16)[name = string("op_93_cast_fp16")];
+            tensor<int32, [1]> target_length_tmp = identity(x = target_length)[name = string("target_length_tmp")];
+        } -> (outputs, output_states_1, output_states_2);
+}

fp16/decoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96cac67a1352e45a2418284779dbd5efe5865159b7b92f328247e5ed31485f52
+size 17876866

fp16/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee08abd75c1af1efc33b494a228e6cfdb25ec762021850a59b9009a7834421b0
+size 14788

fp16/decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:96cac67a1352e45a2418284779dbd5efe5865159b7b92f328247e5ed31485f52
+size 17876866

fp16/decoder.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "716E82BC-EEDE-4F16-85FB-29B774C8BD4D": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "BF97D4BC-AF3A-49CF-AC07-8E15221C4E82": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "BF97D4BC-AF3A-49CF-AC07-8E15221C4E82"
+}

fp16/encoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53e793cb315f6944401df77bdfc332cd08303e72b27f2875394e001f1a4f5898
+size 243

fp16/encoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46d6d38cce3cc26144a2a7cee72b7f41cb7b1b422ac7ce010a17daa7ea2cc981
+size 616

fp16/encoder.mlmodelc/metadata.json

@@ -0,0 +1,155 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Mixed (Float16, Int32)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 1024 × 7)",
+        "shortDescription" : "",
+        "shape" : "[1, 1024, 7]",
+        "name" : "encoded",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "encoded_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 24 × 70 × 1024)",
+        "shortDescription" : "",
+        "shape" : "[1, 24, 70, 1024]",
+        "name" : "cache_channel_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 24 × 1024 × 8)",
+        "shortDescription" : "",
+        "shape" : "[1, 24, 1024, 8]",
+        "name" : "cache_time_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "cache_len_out",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.expandDims" : 5,
+      "Ios18.mul" : 97,
+      "Ios18.softmax" : 24,
+      "Ios18.matmul" : 72,
+      "Ios18.logicalNot" : 2,
+      "Ios18.sigmoid" : 24,
+      "Split" : 24,
+      "Ios18.greaterEqual" : 1,
+      "Select" : 72,
+      "Tile" : 1,
+      "Ios18.add" : 170,
+      "Ios18.layerNorm" : 144,
+      "Ios18.reshape" : 145,
+      "Pad" : 27,
+      "Ios18.linear" : 193,
+      "Ios18.conv" : 77,
+      "Ios18.relu" : 3,
+      "Ios18.concat" : 72,
+      "Ios18.logicalAnd" : 3,
+      "Ios18.silu" : 72,
+      "Ios18.transpose" : 224,
+      "Ios18.cast" : 5,
+      "Stack" : 2,
+      "Ios18.clip" : 1,
+      "Ios18.sliceByIndex" : 147
+    },
+    "computePrecision" : "Mixed (Float16, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "15.0",
+      "tvOS" : "18.0",
+      "visionOS" : "2.0",
+      "watchOS" : "11.0",
+      "iOS" : "18.0",
+      "macCatalyst" : "18.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-03-22",
+      "com.github.apple.coremltools.source" : "torch==2.7.1",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128 × 65)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 65]",
+        "name" : "audio_signal",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 24 × 70 × 1024)",
+        "shortDescription" : "",
+        "shape" : "[1, 24, 70, 1024]",
+        "name" : "cache_last_channel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 24 × 1024 × 8)",
+        "shortDescription" : "",
+        "shape" : "[1, 24, 1024, 8]",
+        "name" : "cache_last_time",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "cache_last_channel_len",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "encoder",
+    "method" : "predict"
+  }
+]

fp16/encoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d39fe49fba8083210400f198545b5e65d8319107f1eb82167896c52d5b491843
+size 1175523648

fp16/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4fe03b263645addcfaa907ea1950a770d5c24a63981bcc96b704cea65d71bb8
+size 698056

fp16/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d39fe49fba8083210400f198545b5e65d8319107f1eb82167896c52d5b491843
+size 1175523648

fp16/encoder.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "46E9C84C-1261-481E-995D-D1CC48828B73": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "4C4FF97E-3218-47F5-89B3-4A9A6CB8FFCD": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "46E9C84C-1261-481E-995D-D1CC48828B73"
+}

fp16/metadata.json

@@ -0,0 +1,25 @@
+{
+  "model": "nvidia/nemotron-speech-streaming-en-0.6b",
+  "sample_rate": 16000,
+  "mel_features": 128,
+  "chunk_mel_frames": 56,
+  "pre_encode_cache": 9,
+  "total_mel_frames": 65,
+  "vocab_size": 1024,
+  "blank_idx": 1024,
+  "encoder_dim": 1024,
+  "decoder_hidden": 640,
+  "decoder_layers": 2,
+  "cache_channel_shape": [
+    24,
+    1,
+    70,
+    1024
+  ],
+  "cache_time_shape": [
+    24,
+    1,
+    1024,
+    8
+  ]
+}