initial commit

.gitignore

@@ -0,0 +1 @@
+.DS_Store

LICENSE_NOTICE.txt

@@ -0,0 +1,7 @@
+Argmax proprietary and confidential. Under NDA.
+
+Copyright 2024 Argmax, Inc. All rights reserved.
+
+Unauthorized access, copying, use, distribution, and or commercialization of this file, via any medium or means is strictly prohibited.
+
+Please contact Argmax for licensing information at [email protected].

README.md

@@ -1,5 +1,24 @@
----
-license: other
-license_name: argmax-fmod-license
-license_link: LICENSE
----
+---
+license: other
+license_name: argmax-fmod-license
+license_link: https://huggingface.co/argmaxinc/whisperkit-pro/blob/main/LICENSE_NOTICE.txt
+pretty_name: "WhisperKit"
+viewer: false
+library_name: whisperkit
+tags:
+- whisper
+- whisperkit
+- coreml
+- asr
+- quantized
+- automatic-speech-recognition
+extra_gated_heading: "WhisperKit Pro is now in early access!"
+extra_gated_description: "WhisperKit Pro is the commercial tier of [WhisperKit](https://github.com/argmaxinc/WhisperKit). Please submit your information below to request early access or directly send an email to [[email protected]](mailto:[email protected])"
+extra_gated_fields:
+  Company: text
+  Work email: text
+  I acknowledge the license notice: checkbox
+extra_gated_button_content: "Submit"
+---
+
+# WhisperKit Pro

config.json

@@ -0,0 +1,112 @@
+{
+    "name": "whisperkit-coreml",
+    "version": "0.2",
+    "device_support": [
+        {
+            "identifiers": ["iPhone11", "iPhone12", "Watch7", "Watch8"],
+            "models": {
+                "default": "openai_whisper-tiny",
+                "supported": [
+                    "openai_whisper-tiny",
+                    "openai_whisper-tiny.en",
+                    "openai_whisper-base",
+                    "openai_whisper-base.en"
+                ]
+            }
+        },
+        {
+            "identifiers": ["iPhone13", "iPad13,18", "iPad13,1"],
+            "models": {
+                "default": "openai_whisper-base",
+                "supported": [
+                    "openai_whisper-tiny",
+                    "openai_whisper-tiny.en",
+                    "openai_whisper-base",
+                    "openai_whisper-base.en",
+                    "openai_whisper-small",
+                    "openai_whisper-small.en"
+                ]
+            }
+        },
+        {
+            "identifiers": [
+                "iPhone14",
+                "iPhone15",
+                "iPhone16",
+                "iPhone17",
+                "iPad14,1",
+                "iPad14,2"
+            ],
+            "models": {
+                "default": "openai_whisper-base",
+                "supported": [
+                    "openai_whisper-tiny",
+                    "openai_whisper-tiny.en",
+                    "openai_whisper-base",
+                    "openai_whisper-base.en",
+                    "openai_whisper-small",
+                    "openai_whisper-small.en",
+                ]
+            }
+        },
+        {
+            "identifiers": [
+                "Mac13",
+                "iMac21",
+                "MacBookAir10,1",
+                "MacBookPro17",
+                "MacBookPro18",
+                "Macmini9",
+                "iPad13,16",
+                "iPad13,4",
+                "iPad13,8"
+            ],
+            "models": {
+                "default": "openai_whisper-large-v3-v20240930",
+                "supported": [
+                    "openai_whisper-tiny",
+                    "openai_whisper-tiny.en",
+                    "openai_whisper-base",
+                    "openai_whisper-base.en",
+                    "openai_whisper-small",
+                    "openai_whisper-small.en",
+                    "distil-whisper_distil-large-v3",
+                    "openai_whisper-large-v3-v20240930",
+                ]
+            }
+        },
+        {
+            "identifiers": [
+                "Mac14",
+                "Mac15",
+                "Mac16",
+                "iPad14,3",
+                "iPad14,4",
+                "iPad14,5",
+                "iPad14,6",
+                "iPad14,8",
+                "iPad14,9",
+                "iPad14,10",
+                "iPad14,11",
+                "iPad16"
+            ],
+            "models": {
+                "default": "openai_whisper-large-v3-v20240930",
+                "supported": [
+                    "openai_whisper-tiny",
+                    "openai_whisper-tiny.en",
+                    "openai_whisper-base",
+                    "openai_whisper-base.en",
+                    "openai_whisper-small",
+                    "openai_whisper-small.en",
+                    "distil-whisper_distil-large-v3",
+                    "distil-whisper_distil-large-v3_turbo",
+                    "openai_whisper-large-v3-v20240930",
+                    "openai_whisper-large-v3-v20240930_turbo",
+                ]
+            }
+        }
+    ],
+    "model_checksums": {
+    }
+}

distil-whisper_distil-large-v3/AudioEncoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1e272b852871322d9a085dfa95707110046e749f525a0db0d8f277709fcde3f
+size 243

distil-whisper_distil-large-v3/AudioEncoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:49b36ed70ce2688383176bec4bd74a1a7716b505ddffd95d5f8eda4063ce6ae3
+size 434

distil-whisper_distil-large-v3/AudioEncoder.mlmodelc/metadata.json

@@ -0,0 +1,90 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 1280 × 1 × 1500)",
+        "shortDescription" : "",
+        "shape" : "[1, 1280, 1, 1500]",
+        "name" : "encoder_output_embeds",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 2 × 1280 × 1 × 1536)",
+        "shortDescription" : "",
+        "shape" : "[2, 1280, 1, 1536]",
+        "name" : "encoder_attn_key_cache",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 2 × 1280 × 1 × 1536)",
+        "shortDescription" : "",
+        "shape" : "[2, 1280, 1, 1536]",
+        "name" : "encoder_attn_value_cache",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.reshape" : 128,
+      "Ios18.batchNorm" : 65,
+      "Ios18.softmax" : 32,
+      "Pad" : 2,
+      "Ios18.concat" : 2,
+      "Ios18.gelu" : 34,
+      "Ios18.layerNorm" : 65,
+      "Ios18.matmul" : 64,
+      "Ios18.conv" : 198,
+      "Ios18.mul" : 32,
+      "Ios18.add" : 65
+    },
+    "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.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.version" : "8.0",
+      "com.github.apple.coremltools.source" : "torch==2.5.1"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 128 × 1 × 3000)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 1, 3000]",
+        "name" : "melspectrogram_features",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "AudioEncoderStateful",
+    "method" : "predict"
+  }
+]

distil-whisper_distil-large-v3/AudioEncoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b43a5d9e21e95067e0af8cf4b8fcbd16cc8e6f99993084f5e67cdf81bde16e79
+size 1287087104

distil-whisper_distil-large-v3/LICENSE_NOTICE.txt

@@ -0,0 +1,7 @@
+Argmax proprietary and confidential. Under NDA.
+
+Copyright 2024 Argmax, Inc. All rights reserved.
+
+Unauthorized access, copying, use, distribution, and or commercialization of this file, via any medium or means is strictly prohibited.
+
+Please contact Argmax for licensing information at [email protected].

distil-whisper_distil-large-v3/MelSpectrogram.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0980462db89a546e1e90888ea38e0a5ddf1f1fec84608802cdbb12f8a5cc7215
+size 243

distil-whisper_distil-large-v3/MelSpectrogram.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:451a5796f1dafb1969fef7bac32cd7fcf51fc763d1e1826ee6211dd046ede15a
+size 329

distil-whisper_distil-large-v3/MelSpectrogram.mlmodelc/metadata.json

@@ -0,0 +1,74 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 128 × 1 × 3000)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 1, 3000]",
+        "name" : "melspectrogram_features",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.mul" : 2,
+      "Ios18.square" : 2,
+      "Ios18.conv" : 2,
+      "Ios18.matmul" : 1,
+      "Ios18.expandDims" : 4,
+      "Ios18.sub" : 1,
+      "Ios18.log" : 1,
+      "Ios18.add" : 3,
+      "Ios18.sliceByIndex" : 1,
+      "Ios18.maximum" : 1,
+      "Ios18.squeeze" : 2,
+      "Ios18.reshape" : 2,
+      "Ios16.reduceMax" : 1,
+      "Identity" : 1,
+      "Pad" : 1
+    },
+    "computePrecision" : "Mixed (Float16, Float32, 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.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.5.1",
+      "com.github.apple.coremltools.version" : "8.0"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 480000)",
+        "shortDescription" : "",
+        "shape" : "[480000]",
+        "name" : "audio",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "MelSpectrogram",
+    "method" : "predict"
+  }
+]

distil-whisper_distil-large-v3/MelSpectrogram.mlmodelc/model.mil

@@ -0,0 +1,66 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3401.3.1"}, {"coremlc-version", "3401.4.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.0"}})]
+{
+    func main<ios18>(tensor<fp16, [480000]> audio) {
+            tensor<int32, [3]> var_10 = const()[name = string("op_10"), val = tensor<int32, [3]>([1, 1, 480000])];
+            tensor<fp16, [1, 1, 480000]> input_1_cast_fp16 = reshape(shape = var_10, x = audio)[name = string("input_1_cast_fp16")];
+            tensor<int32, [6]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 200, 200])];
+            string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("reflect")];
+            fp16 const_1_to_fp16 = const()[name = string("const_1_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 480400]> input_3_cast_fp16 = pad(constant_val = const_1_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")];
+            tensor<int32, [1]> var_22 = const()[name = string("op_22"), val = tensor<int32, [1]>([480400])];
+            tensor<fp16, [480400]> input_cast_fp16 = reshape(shape = var_22, x = input_3_cast_fp16)[name = string("input_cast_fp16")];
+            tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = string("expand_dims_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 480400]> expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = input_cast_fp16)[name = string("expand_dims_0_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, 480400]> expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = expand_dims_0_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, [201, 1, 400]> expand_dims_1_to_fp16 = const()[name = string("expand_dims_1_to_fp16"), val = tensor<fp16, [201, 1, 400]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 201, 3001]> 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, [201, 1, 400]> expand_dims_2_to_fp16 = const()[name = string("expand_dims_2_to_fp16"), val = tensor<fp16, [201, 1, 400]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160960)))];
+            tensor<fp16, [1, 201, 3001]> 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")];
+            tensor<int32, [1]> squeeze_0_axes_0 = const()[name = string("squeeze_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [201, 3001]> squeeze_0_cast_fp16 = squeeze(axes = squeeze_0_axes_0, x = conv_0_cast_fp16)[name = string("squeeze_0_cast_fp16")];
+            tensor<int32, [1]> squeeze_1_axes_0 = const()[name = string("squeeze_1_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [201, 3001]> squeeze_1_cast_fp16 = squeeze(axes = squeeze_1_axes_0, x = conv_1_cast_fp16)[name = string("squeeze_1_cast_fp16")];
+            tensor<fp16, [201, 3001]> square_0_cast_fp16 = square(x = squeeze_0_cast_fp16)[name = string("square_0_cast_fp16")];
+            tensor<fp16, [201, 3001]> square_1_cast_fp16 = square(x = squeeze_1_cast_fp16)[name = string("square_1_cast_fp16")];
+            tensor<fp16, [201, 3001]> add_1_cast_fp16 = add(x = square_0_cast_fp16, y = square_1_cast_fp16)[name = string("add_1_cast_fp16")];
+            tensor<fp16, [201, 3001]> magnitudes_1_cast_fp16 = identity(x = add_1_cast_fp16)[name = string("magnitudes_1_cast_fp16")];
+            tensor<int32, [2]> magnitudes_begin_0 = const()[name = string("magnitudes_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> magnitudes_end_0 = const()[name = string("magnitudes_end_0"), val = tensor<int32, [2]>([201, 3000])];
+            tensor<bool, [2]> magnitudes_end_mask_0 = const()[name = string("magnitudes_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp16, [201, 3000]> magnitudes_cast_fp16 = slice_by_index(begin = magnitudes_begin_0, end = magnitudes_end_0, end_mask = magnitudes_end_mask_0, x = magnitudes_1_cast_fp16)[name = string("magnitudes_cast_fp16")];
+            bool mel_spec_1_transpose_x_0 = const()[name = string("mel_spec_1_transpose_x_0"), val = bool(false)];
+            bool mel_spec_1_transpose_y_0 = const()[name = string("mel_spec_1_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [128, 201]> mel_filters_to_fp16 = const()[name = string("mel_filters_to_fp16"), val = tensor<fp16, [128, 201]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321856)))];
+            tensor<fp16, [128, 3000]> mel_spec_1_cast_fp16 = matmul(transpose_x = mel_spec_1_transpose_x_0, transpose_y = mel_spec_1_transpose_y_0, x = mel_filters_to_fp16, y = magnitudes_cast_fp16)[name = string("mel_spec_1_cast_fp16")];
+            fp16 var_41_to_fp16 = const()[name = string("op_41_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [128, 3000]> mel_spec_cast_fp16 = add(x = mel_spec_1_cast_fp16, y = var_41_to_fp16)[name = string("mel_spec_cast_fp16")];
+            fp32 log_0_epsilon_0 = const()[name = string("log_0_epsilon_0"), val = fp32(0x1p-149)];
+            tensor<fp16, [128, 3000]> log_0_cast_fp16 = log(epsilon = log_0_epsilon_0, x = mel_spec_cast_fp16)[name = string("log_0_cast_fp16")];
+            fp16 mul_0_y_0_to_fp16 = const()[name = string("mul_0_y_0_to_fp16"), val = fp16(0x1.bccp-2)];
+            tensor<fp16, [128, 3000]> mul_0_cast_fp16 = mul(x = log_0_cast_fp16, y = mul_0_y_0_to_fp16)[name = string("mul_0_cast_fp16")];
+            bool var_44_keep_dims_0 = const()[name = string("op_44_keep_dims_0"), val = bool(false)];
+            fp16 var_44_cast_fp16 = reduce_max(keep_dims = var_44_keep_dims_0, x = mul_0_cast_fp16)[name = string("op_44_cast_fp16")];
+            fp16 var_46_to_fp16 = const()[name = string("op_46_to_fp16"), val = fp16(0x1p+3)];
+            fp16 var_47_cast_fp16 = sub(x = var_44_cast_fp16, y = var_46_to_fp16)[name = string("op_47_cast_fp16")];
+            tensor<fp16, [128, 3000]> log_spec_3_cast_fp16 = maximum(x = mul_0_cast_fp16, y = var_47_cast_fp16)[name = string("log_spec_3_cast_fp16")];
+            fp16 var_50_to_fp16 = const()[name = string("op_50_to_fp16"), val = fp16(0x1p+2)];
+            tensor<fp16, [128, 3000]> var_51_cast_fp16 = add(x = log_spec_3_cast_fp16, y = var_50_to_fp16)[name = string("op_51_cast_fp16")];
+            fp16 _inversed_log_spec_y_0_to_fp16 = const()[name = string("_inversed_log_spec_y_0_to_fp16"), val = fp16(0x1p-2)];
+            tensor<fp16, [128, 3000]> _inversed_log_spec_cast_fp16 = mul(x = var_51_cast_fp16, y = _inversed_log_spec_y_0_to_fp16)[name = string("_inversed_log_spec_cast_fp16")];
+            tensor<int32, [1]> var_55_axes_0 = const()[name = string("op_55_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 128, 3000]> var_55_cast_fp16 = expand_dims(axes = var_55_axes_0, x = _inversed_log_spec_cast_fp16)[name = string("op_55_cast_fp16")];
+            tensor<int32, [1]> var_62_axes_0 = const()[name = string("op_62_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 128, 1, 3000]> melspectrogram_features = expand_dims(axes = var_62_axes_0, x = var_55_cast_fp16)[name = string("op_62_cast_fp16")];
+        } -> (melspectrogram_features);
+}

distil-whisper_distil-large-v3/MelSpectrogram.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:009d9fb8f6b589accfa08cebf1c712ef07c3405229ce3cfb3a57ee033c9d8a49
+size 373376

distil-whisper_distil-large-v3/TextDecoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77cb1b565a336e7fc01586698e50aa32d9a2a8f1ca5c439172564f4af0515f5d
+size 243

distil-whisper_distil-large-v3/TextDecoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a5e6f62b5ae897c8f846e22cacbe7d4f7d6bdbeb5f46366e2387f1082676b62
+size 754

distil-whisper_distil-large-v3/TextDecoder.mlmodelc/metadata.json

@@ -0,0 +1,183 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 1 × 51866)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 51866]",
+        "name" : "logits",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 2560 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 2560, 1, 1]",
+        "name" : "key_cache_updates",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 2560 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 2560, 1, 1]",
+        "name" : "value_cache_updates",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 1536)",
+        "shortDescription" : "",
+        "shape" : "[1, 1536]",
+        "name" : "alignment_heads_weights",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.expandDims" : 8,
+      "Ios18.softmax" : 4,
+      "Ios18.mul" : 8,
+      "Ios18.matmul" : 8,
+      "Ios18.batchNorm" : 7,
+      "Ios16.reduceMean" : 1,
+      "Split" : 2,
+      "Ios18.readState" : 5,
+      "Ios18.gather" : 2,
+      "Ios18.add" : 15,
+      "Ios18.layerNorm" : 7,
+      "Ios18.reshape" : 16,
+      "Ios18.linear" : 1,
+      "Ios18.conv" : 16,
+      "Ios18.gelu" : 2,
+      "Ios18.concat" : 3,
+      "Ios18.cast" : 1,
+      "Ios18.transpose" : 1,
+      "Ios18.sliceByIndex" : 44,
+      "Ios18.squeeze" : 1
+    },
+    "computePrecision" : "Mixed (Float16, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 1 × 1536)",
+        "shortDescription" : "",
+        "shape" : "[1, 1536]",
+        "name" : "encoder_attn_key_padding_mask",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 2 × 1280 × 1 × 1536)",
+        "shortDescription" : "",
+        "shape" : "[2, 1280, 1, 1536]",
+        "name" : "encoder_attn_key_cache",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 2 × 1280 × 1 × 1536)",
+        "shortDescription" : "",
+        "shape" : "[2, 1280, 1, 1536]",
+        "name" : "encoder_attn_value_cache",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 2 × 1280 × 1 × 448)",
+        "shortDescription" : "",
+        "shape" : "[2, 1280, 1, 448]",
+        "name" : "self_attn_key_cache",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 2 × 1280 × 1 × 448)",
+        "shortDescription" : "",
+        "shape" : "[2, 1280, 1, 448]",
+        "name" : "self_attn_value_cache",
+        "type" : "State"
+      }
+    ],
+    "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.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.5.1",
+      "com.github.apple.coremltools.version" : "8.0"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "input_ids",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "cache_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 448)",
+        "shortDescription" : "",
+        "shape" : "[1, 448]",
+        "name" : "kv_cache_update_mask",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 448)",
+        "shortDescription" : "",
+        "shape" : "[1, 448]",
+        "name" : "decoder_key_padding_mask",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "TextDecoderStateful",
+    "method" : "predict"
+  }
+]

distil-whisper_distil-large-v3/TextDecoder.mlmodelc/model.mil

@@ -0,0 +1,529 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3401.3.1"}, {"coremlc-version", "3401.4.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.0"}})]
+{
+    func main<ios18>(tensor<int32, [1]> cache_length, tensor<fp16, [1, 448]> decoder_key_padding_mask, state<tensor<fp16, [2, 1280, 1, 1536]>> encoder_attn_key_cache, state<tensor<fp16, [1, 1536]>> encoder_attn_key_padding_mask, state<tensor<fp16, [2, 1280, 1, 1536]>> encoder_attn_value_cache, tensor<int32, [1]> input_ids, tensor<fp16, [1, 448]> kv_cache_update_mask, state<tensor<fp16, [2, 1280, 1, 448]>> self_attn_key_cache, state<tensor<fp16, [2, 1280, 1, 448]>> self_attn_value_cache) {
+            int32 var_22_axis_0 = const()[name = string("op_22_axis_0"), val = int32(0)];
+            int32 var_22_batch_dims_0 = const()[name = string("op_22_batch_dims_0"), val = int32(0)];
+            bool var_22_validate_indices_0 = const()[name = string("op_22_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [51866, 1280]> embed_tokens_weight_to_fp16 = const()[name = string("embed_tokens_weight_to_fp16"), val = tensor<fp16, [51866, 1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 1280]> var_22_cast_fp16 = gather(axis = var_22_axis_0, batch_dims = var_22_batch_dims_0, indices = input_ids, validate_indices = var_22_validate_indices_0, x = embed_tokens_weight_to_fp16)[name = string("op_22_cast_fp16")];
+            int32 var_26_axis_0 = const()[name = string("op_26_axis_0"), val = int32(0)];
+            int32 var_26_batch_dims_0 = const()[name = string("op_26_batch_dims_0"), val = int32(0)];
+            bool var_26_validate_indices_0 = const()[name = string("op_26_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [448, 1280]> embed_positions_weight_to_fp16 = const()[name = string("embed_positions_weight_to_fp16"), val = tensor<fp16, [448, 1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132777088)))];
+            string cache_length_to_uint16_dtype_0 = const()[name = string("cache_length_to_uint16_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1]> cache_length_to_uint16 = cast(dtype = cache_length_to_uint16_dtype_0, x = cache_length)[name = string("cast_43")];
+            tensor<fp16, [1, 1280]> var_26_cast_fp16_cast_uint16 = gather(axis = var_26_axis_0, batch_dims = var_26_batch_dims_0, indices = cache_length_to_uint16, validate_indices = var_26_validate_indices_0, x = embed_positions_weight_to_fp16)[name = string("op_26_cast_fp16_cast_uint16")];
+            tensor<fp16, [1, 1280]> hidden_states_1_cast_fp16 = add(x = var_22_cast_fp16, y = var_26_cast_fp16_cast_uint16)[name = string("hidden_states_1_cast_fp16")];
+            tensor<int32, [1]> var_40_axes_0 = const()[name = string("op_40_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1280, 1]> var_40_cast_fp16 = expand_dims(axes = var_40_axes_0, x = hidden_states_1_cast_fp16)[name = string("op_40_cast_fp16")];
+            tensor<int32, [1]> inputs_1_axes_0 = const()[name = string("inputs_1_axes_0"), val = tensor<int32, [1]>([3])];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_40_cast_fp16)[name = string("inputs_1_cast_fp16")];
+            tensor<fp16, [2, 1280, 1, 448]> read_state_0 = read_state(input = self_attn_key_cache)[name = string("read_state_0")];
+            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_45_axis_0 = const()[name = string("op_45_axis_0"), val = int32(0)];
+            tensor<fp16, [1, 1280, 1, 448]> var_45_cast_fp16_0, tensor<fp16, [1, 1280, 1, 448]> var_45_cast_fp16_1 = split(axis = var_45_axis_0, split_sizes = tile_0, x = read_state_0)[name = string("op_45_cast_fp16")];
+            tensor<fp16, [2, 1280, 1, 448]> read_state_1 = read_state(input = self_attn_value_cache)[name = string("read_state_1")];
+            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_50_axis_0 = const()[name = string("op_50_axis_0"), val = int32(0)];
+            tensor<fp16, [1, 1280, 1, 448]> var_50_cast_fp16_0, tensor<fp16, [1, 1280, 1, 448]> var_50_cast_fp16_1 = split(axis = var_50_axis_0, split_sizes = tile_1, x = read_state_1)[name = string("op_50_cast_fp16")];
+            tensor<fp16, [2, 1280, 1, 1536]> read_state_2 = read_state(input = encoder_attn_key_cache)[name = string("read_state_2")];
+            tensor<int32, [4]> obj_17_begin_0 = const()[name = string("obj_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> obj_17_end_0 = const()[name = string("obj_17_end_0"), val = tensor<int32, [4]>([1, 1280, 1, 1536])];
+            tensor<bool, [4]> obj_17_end_mask_0 = const()[name = string("obj_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 1280, 1, 1536]> obj_17_cast_fp16 = slice_by_index(begin = obj_17_begin_0, end = obj_17_end_0, end_mask = obj_17_end_mask_0, x = read_state_2)[name = string("obj_17_cast_fp16")];
+            tensor<fp16, [2, 1280, 1, 1536]> read_state_3 = read_state(input = encoder_attn_value_cache)[name = string("read_state_3")];
+            tensor<int32, [4]> obj_19_begin_0 = const()[name = string("obj_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> obj_19_end_0 = const()[name = string("obj_19_end_0"), val = tensor<int32, [4]>([1, 1280, 1, 1536])];
+            tensor<bool, [4]> obj_19_end_mask_0 = const()[name = string("obj_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 1280, 1, 1536]> obj_19_cast_fp16 = slice_by_index(begin = obj_19_begin_0, end = obj_19_end_0, end_mask = obj_19_end_mask_0, x = read_state_3)[name = string("obj_19_cast_fp16")];
+            int32 var_68 = const()[name = string("op_68"), val = int32(3)];
+            tensor<int32, [1]> out_1_axes_0 = const()[name = string("out_1_axes_0"), val = tensor<int32, [1]>([1])];
+            fp16 var_93_to_fp16 = const()[name = string("op_93_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_93_to_fp16, x = inputs_1_cast_fp16)[name = string("out_1_cast_fp16")];
+            tensor<fp16, [1280]> obj_5_mean_0_to_fp16 = const()[name = string("obj_5_mean_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133924032)))];
+            tensor<fp16, [1280]> obj_5_variance_0_to_fp16 = const()[name = string("obj_5_variance_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133926656)))];
+            tensor<fp16, [1280]> obj_5_gamma_0_to_fp16 = const()[name = string("obj_5_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133929280)))];
+            tensor<fp16, [1280]> obj_5_beta_0_to_fp16 = const()[name = string("obj_5_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133931904)))];
+            fp16 obj_5_epsilon_0_to_fp16 = const()[name = string("obj_5_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> obj_5_cast_fp16 = batch_norm(beta = obj_5_beta_0_to_fp16, epsilon = obj_5_epsilon_0_to_fp16, gamma = obj_5_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_1_cast_fp16)[name = string("obj_5_cast_fp16")];
+            string query_1_pad_type_0 = const()[name = string("query_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> query_1_strides_0 = const()[name = string("query_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> query_1_pad_0 = const()[name = string("query_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> query_1_dilations_0 = const()[name = string("query_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 query_1_groups_0 = const()[name = string("query_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133934528)))];
+            tensor<fp16, [1280]> layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137211392)))];
+            tensor<fp16, [1, 1280, 1, 1]> query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = query_1_dilations_0, groups = query_1_groups_0, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = query_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = string("query_1_cast_fp16")];
+            string current_key_1_pad_type_0 = const()[name = string("current_key_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> current_key_1_strides_0 = const()[name = string("current_key_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> current_key_1_pad_0 = const()[name = string("current_key_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> current_key_1_dilations_0 = const()[name = string("current_key_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 current_key_1_groups_0 = const()[name = string("current_key_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137214016)))];
+            tensor<fp16, [1, 1280, 1, 1]> current_key_1_cast_fp16 = conv(dilations = current_key_1_dilations_0, groups = current_key_1_groups_0, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = current_key_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = string("current_key_1_cast_fp16")];
+            string current_value_1_pad_type_0 = const()[name = string("current_value_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> current_value_1_strides_0 = const()[name = string("current_value_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> current_value_1_pad_0 = const()[name = string("current_value_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> current_value_1_dilations_0 = const()[name = string("current_value_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 current_value_1_groups_0 = const()[name = string("current_value_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140490880)))];
+            tensor<fp16, [1280]> layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143767744)))];
+            tensor<fp16, [1, 1280, 1, 1]> current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = current_value_1_dilations_0, groups = current_value_1_groups_0, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = current_value_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = string("current_value_1_cast_fp16")];
+            tensor<int32, [1]> var_128_axes_0 = const()[name = string("op_128_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 448]> var_128_cast_fp16 = expand_dims(axes = var_128_axes_0, x = kv_cache_update_mask)[name = string("op_128_cast_fp16")];
+            tensor<int32, [1]> var_129_axes_0 = const()[name = string("op_129_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 448]> var_129_cast_fp16 = expand_dims(axes = var_129_axes_0, x = var_128_cast_fp16)[name = string("op_129_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 448]> var_131_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_129_cast_fp16)[name = string("op_131_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 448]> key_1_cast_fp16 = add(x = var_45_cast_fp16_0, y = var_131_cast_fp16)[name = string("key_1_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 448]> var_133_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_129_cast_fp16)[name = string("op_133_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 448]> value_1_cast_fp16 = add(x = var_50_cast_fp16_0, y = var_133_cast_fp16)[name = string("value_1_cast_fp16")];
+            tensor<int32, [4]> var_136 = const()[name = string("op_136"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1]> mh_q_1_cast_fp16 = reshape(shape = var_136, x = query_1_cast_fp16)[name = string("mh_q_1_cast_fp16")];
+            fp16 var_138_to_fp16 = const()[name = string("op_138_to_fp16"), val = fp16(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1]> var_139_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_138_to_fp16)[name = string("op_139_cast_fp16")];
+            tensor<int32, [4]> var_140 = const()[name = string("op_140"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 448]> var_141_cast_fp16 = reshape(shape = var_140, x = key_1_cast_fp16)[name = string("op_141_cast_fp16")];
+            bool mh_w_1_transpose_x_0 = const()[name = string("mh_w_1_transpose_x_0"), val = bool(true)];
+            bool mh_w_1_transpose_y_0 = const()[name = string("mh_w_1_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 20, 1, 448]> mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_139_cast_fp16, y = var_141_cast_fp16)[name = string("mh_w_1_cast_fp16")];
+            tensor<int32, [1]> var_145_axes_0 = const()[name = string("op_145_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 448]> var_145_cast_fp16 = expand_dims(axes = var_145_axes_0, x = decoder_key_padding_mask)[name = string("op_145_cast_fp16")];
+            tensor<int32, [1]> var_146_axes_0 = const()[name = string("op_146_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 448]> var_146_cast_fp16 = expand_dims(axes = var_146_axes_0, x = var_145_cast_fp16)[name = string("op_146_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 448]> mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_146_cast_fp16)[name = string("mh_w_3_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 448]> var_149_cast_fp16 = softmax(axis = var_68, x = mh_w_3_cast_fp16)[name = string("op_149_cast_fp16")];
+            tensor<int32, [4]> var_150 = const()[name = string("op_150"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 448]> var_151_cast_fp16 = reshape(shape = var_150, x = value_1_cast_fp16)[name = string("op_151_cast_fp16")];
+            bool attn_1_transpose_x_0 = const()[name = string("attn_1_transpose_x_0"), val = bool(false)];
+            bool attn_1_transpose_y_0 = const()[name = string("attn_1_transpose_y_0"), val = bool(true)];
+            tensor<fp16, [1, 20, 64, 1]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_151_cast_fp16, y = var_149_cast_fp16)[name = string("attn_1_cast_fp16")];
+            tensor<int32, [4]> var_154 = const()[name = string("op_154"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1]> input_1_cast_fp16 = reshape(shape = var_154, x = attn_1_cast_fp16)[name = string("input_1_cast_fp16")];
+            string obj_11_pad_type_0 = const()[name = string("obj_11_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> obj_11_strides_0 = const()[name = string("obj_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> obj_11_pad_0 = const()[name = string("obj_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> obj_11_dilations_0 = const()[name = string("obj_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 obj_11_groups_0 = const()[name = string("obj_11_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143770368)))];
+            tensor<fp16, [1280]> layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147047232)))];
+            tensor<fp16, [1, 1280, 1, 1]> obj_11_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = obj_11_dilations_0, groups = obj_11_groups_0, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = obj_11_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("obj_11_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_11_cast_fp16)[name = string("inputs_3_cast_fp16")];
+            tensor<int32, [1]> out_3_axes_0 = const()[name = string("out_3_axes_0"), val = tensor<int32, [1]>([1])];
+            fp16 var_176_to_fp16 = const()[name = string("op_176_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_176_to_fp16, x = inputs_3_cast_fp16)[name = string("out_3_cast_fp16")];
+            tensor<fp16, [1280]> obj_13_gamma_0_to_fp16 = const()[name = string("obj_13_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147049856)))];
+            tensor<fp16, [1280]> obj_13_beta_0_to_fp16 = const()[name = string("obj_13_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147052480)))];
+            fp16 obj_13_epsilon_0_to_fp16 = const()[name = string("obj_13_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> obj_13_cast_fp16 = batch_norm(beta = obj_13_beta_0_to_fp16, epsilon = obj_13_epsilon_0_to_fp16, gamma = obj_13_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_3_cast_fp16)[name = string("obj_13_cast_fp16")];
+            string query_3_pad_type_0 = const()[name = string("query_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> query_3_strides_0 = const()[name = string("query_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> query_3_pad_0 = const()[name = string("query_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> query_3_dilations_0 = const()[name = string("query_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 query_3_groups_0 = const()[name = string("query_3_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147055104)))];
+            tensor<fp16, [1280]> layers_0_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("layers_0_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150331968)))];
+            tensor<fp16, [1, 1280, 1, 1]> query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = query_3_dilations_0, groups = query_3_groups_0, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = query_3_strides_0, weight = layers_0_encoder_attn_q_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = string("query_3_cast_fp16")];
+            tensor<int32, [4]> var_196 = const()[name = string("op_196"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1]> mh_q_3_cast_fp16 = reshape(shape = var_196, x = query_3_cast_fp16)[name = string("mh_q_3_cast_fp16")];
+            fp16 var_198_to_fp16 = const()[name = string("op_198_to_fp16"), val = fp16(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1]> var_199_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_198_to_fp16)[name = string("op_199_cast_fp16")];
+            tensor<int32, [4]> var_200 = const()[name = string("op_200"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1536]> var_201_cast_fp16 = reshape(shape = var_200, x = obj_17_cast_fp16)[name = string("op_201_cast_fp16")];
+            bool mh_w_5_transpose_x_0 = const()[name = string("mh_w_5_transpose_x_0"), val = bool(true)];
+            bool mh_w_5_transpose_y_0 = const()[name = string("mh_w_5_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 20, 1, 1536]> mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_199_cast_fp16, y = var_201_cast_fp16)[name = string("mh_w_5_cast_fp16")];
+            tensor<fp16, [1, 1536]> read_state_4 = read_state(input = encoder_attn_key_padding_mask)[name = string("read_state_4")];
+            tensor<int32, [1]> var_205_axes_0 = const()[name = string("op_205_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1536]> var_205_cast_fp16 = expand_dims(axes = var_205_axes_0, x = read_state_4)[name = string("op_205_cast_fp16")];
+            tensor<int32, [1]> var_206_axes_0 = const()[name = string("op_206_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 1536]> var_206_cast_fp16 = expand_dims(axes = var_206_axes_0, x = var_205_cast_fp16)[name = string("op_206_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 1536]> mh_w_7_cast_fp16 = add(x = mh_w_5_cast_fp16, y = var_206_cast_fp16)[name = string("mh_w_7_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 1536]> obj_23_cast_fp16 = softmax(axis = var_68, x = mh_w_7_cast_fp16)[name = string("obj_23_cast_fp16")];
+            tensor<int32, [4]> var_210 = const()[name = string("op_210"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1536]> var_211_cast_fp16 = reshape(shape = var_210, x = obj_19_cast_fp16)[name = string("op_211_cast_fp16")];
+            bool attn_3_transpose_x_0 = const()[name = string("attn_3_transpose_x_0"), val = bool(false)];
+            bool attn_3_transpose_y_0 = const()[name = string("attn_3_transpose_y_0"), val = bool(true)];
+            tensor<fp16, [1, 20, 64, 1]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_211_cast_fp16, y = obj_23_cast_fp16)[name = string("attn_3_cast_fp16")];
+            tensor<int32, [4]> var_214 = const()[name = string("op_214"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1]> input_3_cast_fp16 = reshape(shape = var_214, x = attn_3_cast_fp16)[name = string("input_3_cast_fp16")];
+            string obj_21_pad_type_0 = const()[name = string("obj_21_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> obj_21_strides_0 = const()[name = string("obj_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> obj_21_pad_0 = const()[name = string("obj_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> obj_21_dilations_0 = const()[name = string("obj_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 obj_21_groups_0 = const()[name = string("obj_21_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150334592)))];
+            tensor<fp16, [1280]> layers_0_encoder_attn_o_proj_bias_to_fp16 = const()[name = string("layers_0_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153611456)))];
+            tensor<fp16, [1, 1280, 1, 1]> obj_21_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = obj_21_dilations_0, groups = obj_21_groups_0, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = obj_21_strides_0, weight = layers_0_encoder_attn_o_proj_weight_to_fp16, x = input_3_cast_fp16)[name = string("obj_21_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_21_cast_fp16)[name = string("inputs_5_cast_fp16")];
+            tensor<int32, [1]> out_5_axes_0 = const()[name = string("out_5_axes_0"), val = tensor<int32, [1]>([1])];
+            fp16 var_232_to_fp16 = const()[name = string("op_232_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_232_to_fp16, x = inputs_5_cast_fp16)[name = string("out_5_cast_fp16")];
+            tensor<fp16, [1280]> input_5_gamma_0_to_fp16 = const()[name = string("input_5_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153614080)))];
+            tensor<fp16, [1280]> input_5_beta_0_to_fp16 = const()[name = string("input_5_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153616704)))];
+            fp16 input_5_epsilon_0_to_fp16 = const()[name = string("input_5_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> input_5_cast_fp16 = batch_norm(beta = input_5_beta_0_to_fp16, epsilon = input_5_epsilon_0_to_fp16, gamma = input_5_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_5_cast_fp16)[name = string("input_5_cast_fp16")];
+            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_0_fc1_weight_to_fp16 = const()[name = string("layers_0_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153619328)))];
+            tensor<fp16, [5120]> layers_0_fc1_bias_to_fp16 = const()[name = string("layers_0_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166726592)))];
+            tensor<fp16, [1, 5120, 1, 1]> input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_0_fc1_weight_to_fp16, x = input_5_cast_fp16)[name = string("input_7_cast_fp16")];
+            string input_9_mode_0 = const()[name = string("input_9_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1]> input_9_cast_fp16 = gelu(mode = input_9_mode_0, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")];
+            string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_0_fc2_weight_to_fp16 = const()[name = string("layers_0_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166736896)))];
+            tensor<fp16, [1280]> layers_0_fc2_bias_to_fp16 = const()[name = string("layers_0_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179844160)))];
+            tensor<fp16, [1, 1280, 1, 1]> hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_fc2_weight_to_fp16, x = input_9_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = string("inputs_7_cast_fp16")];
+            tensor<int32, [4]> obj_35_begin_0 = const()[name = string("obj_35_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> obj_35_end_0 = const()[name = string("obj_35_end_0"), val = tensor<int32, [4]>([2, 1280, 1, 1536])];
+            tensor<bool, [4]> obj_35_end_mask_0 = const()[name = string("obj_35_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 1280, 1, 1536]> obj_35_cast_fp16 = slice_by_index(begin = obj_35_begin_0, end = obj_35_end_0, end_mask = obj_35_end_mask_0, x = read_state_2)[name = string("obj_35_cast_fp16")];
+            tensor<int32, [4]> obj_37_begin_0 = const()[name = string("obj_37_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> obj_37_end_0 = const()[name = string("obj_37_end_0"), val = tensor<int32, [4]>([2, 1280, 1, 1536])];
+            tensor<bool, [4]> obj_37_end_mask_0 = const()[name = string("obj_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 1280, 1, 1536]> obj_37_cast_fp16 = slice_by_index(begin = obj_37_begin_0, end = obj_37_end_0, end_mask = obj_37_end_mask_0, x = read_state_3)[name = string("obj_37_cast_fp16")];
+            int32 var_277 = const()[name = string("op_277"), val = int32(3)];
+            tensor<int32, [1]> out_7_axes_0 = const()[name = string("out_7_axes_0"), val = tensor<int32, [1]>([1])];
+            fp16 var_302_to_fp16 = const()[name = string("op_302_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_302_to_fp16, x = inputs_7_cast_fp16)[name = string("out_7_cast_fp16")];
+            tensor<fp16, [1280]> obj_25_gamma_0_to_fp16 = const()[name = string("obj_25_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179846784)))];
+            tensor<fp16, [1280]> obj_25_beta_0_to_fp16 = const()[name = string("obj_25_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179849408)))];
+            fp16 obj_25_epsilon_0_to_fp16 = const()[name = string("obj_25_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> obj_25_cast_fp16 = batch_norm(beta = obj_25_beta_0_to_fp16, epsilon = obj_25_epsilon_0_to_fp16, gamma = obj_25_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_7_cast_fp16)[name = string("obj_25_cast_fp16")];
+            string query_5_pad_type_0 = const()[name = string("query_5_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> query_5_strides_0 = const()[name = string("query_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> query_5_pad_0 = const()[name = string("query_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> query_5_dilations_0 = const()[name = string("query_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 query_5_groups_0 = const()[name = string("query_5_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179852032)))];
+            tensor<fp16, [1280]> layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183128896)))];
+            tensor<fp16, [1, 1280, 1, 1]> query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = query_5_dilations_0, groups = query_5_groups_0, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = query_5_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = string("query_5_cast_fp16")];
+            string current_key_pad_type_0 = const()[name = string("current_key_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> current_key_strides_0 = const()[name = string("current_key_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> current_key_pad_0 = const()[name = string("current_key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> current_key_dilations_0 = const()[name = string("current_key_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 current_key_groups_0 = const()[name = string("current_key_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183131520)))];
+            tensor<fp16, [1, 1280, 1, 1]> current_key_cast_fp16 = conv(dilations = current_key_dilations_0, groups = current_key_groups_0, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = current_key_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = string("current_key_cast_fp16")];
+            string current_value_pad_type_0 = const()[name = string("current_value_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> current_value_strides_0 = const()[name = string("current_value_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> current_value_pad_0 = const()[name = string("current_value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> current_value_dilations_0 = const()[name = string("current_value_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 current_value_groups_0 = const()[name = string("current_value_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186408384)))];
+            tensor<fp16, [1280]> layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189685248)))];
+            tensor<fp16, [1, 1280, 1, 1]> current_value_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = current_value_dilations_0, groups = current_value_groups_0, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = current_value_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = string("current_value_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 448]> var_340_cast_fp16 = mul(x = current_key_cast_fp16, y = var_129_cast_fp16)[name = string("op_340_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 448]> key_cast_fp16 = add(x = var_45_cast_fp16_1, y = var_340_cast_fp16)[name = string("key_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 448]> var_342_cast_fp16 = mul(x = current_value_cast_fp16, y = var_129_cast_fp16)[name = string("op_342_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 448]> value_cast_fp16 = add(x = var_50_cast_fp16_1, y = var_342_cast_fp16)[name = string("value_cast_fp16")];
+            tensor<int32, [4]> var_345 = const()[name = string("op_345"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1]> mh_q_5_cast_fp16 = reshape(shape = var_345, x = query_5_cast_fp16)[name = string("mh_q_5_cast_fp16")];
+            fp16 var_347_to_fp16 = const()[name = string("op_347_to_fp16"), val = fp16(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1]> var_348_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_347_to_fp16)[name = string("op_348_cast_fp16")];
+            tensor<int32, [4]> var_349 = const()[name = string("op_349"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 448]> var_350_cast_fp16 = reshape(shape = var_349, x = key_cast_fp16)[name = string("op_350_cast_fp16")];
+            bool mh_w_9_transpose_x_0 = const()[name = string("mh_w_9_transpose_x_0"), val = bool(true)];
+            bool mh_w_9_transpose_y_0 = const()[name = string("mh_w_9_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 20, 1, 448]> mh_w_9_cast_fp16 = matmul(transpose_x = mh_w_9_transpose_x_0, transpose_y = mh_w_9_transpose_y_0, x = var_348_cast_fp16, y = var_350_cast_fp16)[name = string("mh_w_9_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 448]> mh_w_11_cast_fp16 = add(x = mh_w_9_cast_fp16, y = var_146_cast_fp16)[name = string("mh_w_11_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 448]> var_358_cast_fp16 = softmax(axis = var_277, x = mh_w_11_cast_fp16)[name = string("op_358_cast_fp16")];
+            tensor<int32, [4]> var_359 = const()[name = string("op_359"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 448]> var_360_cast_fp16 = reshape(shape = var_359, x = value_cast_fp16)[name = string("op_360_cast_fp16")];
+            bool attn_5_transpose_x_0 = const()[name = string("attn_5_transpose_x_0"), val = bool(false)];
+            bool attn_5_transpose_y_0 = const()[name = string("attn_5_transpose_y_0"), val = bool(true)];
+            tensor<fp16, [1, 20, 64, 1]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_360_cast_fp16, y = var_358_cast_fp16)[name = string("attn_5_cast_fp16")];
+            tensor<int32, [4]> var_363 = const()[name = string("op_363"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1]> input_11_cast_fp16 = reshape(shape = var_363, x = attn_5_cast_fp16)[name = string("input_11_cast_fp16")];
+            string obj_31_pad_type_0 = const()[name = string("obj_31_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> obj_31_strides_0 = const()[name = string("obj_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> obj_31_pad_0 = const()[name = string("obj_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> obj_31_dilations_0 = const()[name = string("obj_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 obj_31_groups_0 = const()[name = string("obj_31_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189687872)))];
+            tensor<fp16, [1280]> layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192964736)))];
+            tensor<fp16, [1, 1280, 1, 1]> obj_31_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = obj_31_dilations_0, groups = obj_31_groups_0, pad = obj_31_pad_0, pad_type = obj_31_pad_type_0, strides = obj_31_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = string("obj_31_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_31_cast_fp16)[name = string("inputs_9_cast_fp16")];
+            tensor<int32, [1]> out_9_axes_0 = const()[name = string("out_9_axes_0"), val = tensor<int32, [1]>([1])];
+            fp16 var_385_to_fp16 = const()[name = string("op_385_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_385_to_fp16, x = inputs_9_cast_fp16)[name = string("out_9_cast_fp16")];
+            tensor<fp16, [1280]> obj_33_gamma_0_to_fp16 = const()[name = string("obj_33_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192967360)))];
+            tensor<fp16, [1280]> obj_33_beta_0_to_fp16 = const()[name = string("obj_33_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192969984)))];
+            fp16 obj_33_epsilon_0_to_fp16 = const()[name = string("obj_33_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> obj_33_cast_fp16 = batch_norm(beta = obj_33_beta_0_to_fp16, epsilon = obj_33_epsilon_0_to_fp16, gamma = obj_33_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_9_cast_fp16)[name = string("obj_33_cast_fp16")];
+            string query_pad_type_0 = const()[name = string("query_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> query_strides_0 = const()[name = string("query_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> query_pad_0 = const()[name = string("query_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> query_dilations_0 = const()[name = string("query_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 query_groups_0 = const()[name = string("query_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192972608)))];
+            tensor<fp16, [1280]> layers_1_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("layers_1_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196249472)))];
+            tensor<fp16, [1, 1280, 1, 1]> query_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = query_dilations_0, groups = query_groups_0, pad = query_pad_0, pad_type = query_pad_type_0, strides = query_strides_0, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_33_cast_fp16)[name = string("query_cast_fp16")];
+            tensor<int32, [4]> var_405 = const()[name = string("op_405"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1]> mh_q_cast_fp16 = reshape(shape = var_405, x = query_cast_fp16)[name = string("mh_q_cast_fp16")];
+            fp16 var_407_to_fp16 = const()[name = string("op_407_to_fp16"), val = fp16(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1]> var_408_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_407_to_fp16)[name = string("op_408_cast_fp16")];
+            tensor<int32, [4]> var_409 = const()[name = string("op_409"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1536]> var_410_cast_fp16 = reshape(shape = var_409, x = obj_35_cast_fp16)[name = string("op_410_cast_fp16")];
+            bool mh_w_13_transpose_x_0 = const()[name = string("mh_w_13_transpose_x_0"), val = bool(true)];
+            bool mh_w_13_transpose_y_0 = const()[name = string("mh_w_13_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 20, 1, 1536]> mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_408_cast_fp16, y = var_410_cast_fp16)[name = string("mh_w_13_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 1536]> mh_w_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_206_cast_fp16)[name = string("mh_w_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 1536]> obj_41_cast_fp16 = softmax(axis = var_277, x = mh_w_cast_fp16)[name = string("obj_41_cast_fp16")];
+            tensor<int32, [4]> var_419 = const()[name = string("op_419"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1536]> var_420_cast_fp16 = reshape(shape = var_419, x = obj_37_cast_fp16)[name = string("op_420_cast_fp16")];
+            bool attn_transpose_x_0 = const()[name = string("attn_transpose_x_0"), val = bool(false)];
+            bool attn_transpose_y_0 = const()[name = string("attn_transpose_y_0"), val = bool(true)];
+            tensor<fp16, [1, 20, 64, 1]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_420_cast_fp16, y = obj_41_cast_fp16)[name = string("attn_cast_fp16")];
+            tensor<int32, [4]> var_423 = const()[name = string("op_423"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1]> input_13_cast_fp16 = reshape(shape = var_423, x = attn_cast_fp16)[name = string("input_13_cast_fp16")];
+            string obj_39_pad_type_0 = const()[name = string("obj_39_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> obj_39_strides_0 = const()[name = string("obj_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> obj_39_pad_0 = const()[name = string("obj_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> obj_39_dilations_0 = const()[name = string("obj_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 obj_39_groups_0 = const()[name = string("obj_39_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196252096)))];
+            tensor<fp16, [1280]> layers_1_encoder_attn_o_proj_bias_to_fp16 = const()[name = string("layers_1_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199528960)))];
+            tensor<fp16, [1, 1280, 1, 1]> obj_39_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = obj_39_dilations_0, groups = obj_39_groups_0, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = obj_39_strides_0, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = string("obj_39_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_39_cast_fp16)[name = string("inputs_11_cast_fp16")];
+            tensor<int32, [1]> out_11_axes_0 = const()[name = string("out_11_axes_0"), val = tensor<int32, [1]>([1])];
+            fp16 var_444_to_fp16 = const()[name = string("op_444_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_444_to_fp16, x = inputs_11_cast_fp16)[name = string("out_11_cast_fp16")];
+            tensor<fp16, [1280]> input_15_gamma_0_to_fp16 = const()[name = string("input_15_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199531584)))];
+            tensor<fp16, [1280]> input_15_beta_0_to_fp16 = const()[name = string("input_15_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199534208)))];
+            fp16 input_15_epsilon_0_to_fp16 = const()[name = string("input_15_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> input_15_cast_fp16 = batch_norm(beta = input_15_beta_0_to_fp16, epsilon = input_15_epsilon_0_to_fp16, gamma = input_15_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_11_cast_fp16)[name = string("input_15_cast_fp16")];
+            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_1_fc1_weight_to_fp16 = const()[name = string("layers_1_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199536832)))];
+            tensor<fp16, [5120]> layers_1_fc1_bias_to_fp16 = const()[name = string("layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212644096)))];
+            tensor<fp16, [1, 5120, 1, 1]> input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")];
+            string input_mode_0 = const()[name = string("input_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1]> input_cast_fp16 = gelu(mode = input_mode_0, x = input_17_cast_fp16)[name = string("input_cast_fp16")];
+            string hidden_states_5_pad_type_0 = const()[name = string("hidden_states_5_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> hidden_states_5_strides_0 = const()[name = string("hidden_states_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> hidden_states_5_pad_0 = const()[name = string("hidden_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> hidden_states_5_dilations_0 = const()[name = string("hidden_states_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 hidden_states_5_groups_0 = const()[name = string("hidden_states_5_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_1_fc2_weight_to_fp16 = const()[name = string("layers_1_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212654400)))];
+            tensor<fp16, [1280]> layers_1_fc2_bias_to_fp16 = const()[name = string("layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225761664)))];
+            tensor<fp16, [1, 1280, 1, 1]> hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = hidden_states_5_dilations_0, groups = hidden_states_5_groups_0, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = hidden_states_5_strides_0, weight = layers_1_fc2_weight_to_fp16, x = input_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = string("inputs_cast_fp16")];
+            tensor<int32, [1]> out_axes_0 = const()[name = string("out_axes_0"), val = tensor<int32, [1]>([1])];
+            fp16 var_487_to_fp16 = const()[name = string("op_487_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_487_to_fp16, x = inputs_cast_fp16)[name = string("out_cast_fp16")];
+            tensor<fp16, [1280]> hidden_states_gamma_0_to_fp16 = const()[name = string("hidden_states_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225764288)))];
+            tensor<fp16, [1280]> hidden_states_beta_0_to_fp16 = const()[name = string("hidden_states_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225766912)))];
+            fp16 hidden_states_epsilon_0_to_fp16 = const()[name = string("hidden_states_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_cast_fp16)[name = string("hidden_states_cast_fp16")];
+            tensor<int32, [1]> var_498_axes_0 = const()[name = string("op_498_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1280, 1]> var_498_cast_fp16 = squeeze(axes = var_498_axes_0, x = hidden_states_cast_fp16)[name = string("op_498_cast_fp16")];
+            tensor<int32, [3]> var_501_perm_0 = const()[name = string("op_501_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [51866]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [51866]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225769536)))];
+            tensor<fp16, [1, 1, 1280]> var_501_cast_fp16 = transpose(perm = var_501_perm_0, x = var_498_cast_fp16)[name = string("transpose_0")];
+            tensor<fp16, [1, 1, 51866]> logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_501_cast_fp16)[name = string("linear_0_cast_fp16")];
+            int32 var_505 = const()[name = string("op_505"), val = int32(1)];
+            bool obj_45_interleave_0 = const()[name = string("obj_45_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2560, 1, 1]> key_cache_updates = concat(axis = var_505, interleave = obj_45_interleave_0, values = (current_key_1_cast_fp16, current_key_cast_fp16))[name = string("obj_45_cast_fp16")];
+            int32 var_508 = const()[name = string("op_508"), val = int32(1)];
+            bool obj_47_interleave_0 = const()[name = string("obj_47_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2560, 1, 1]> value_cache_updates = concat(axis = var_508, interleave = obj_47_interleave_0, values = (current_value_1_cast_fp16, current_value_cast_fp16))[name = string("obj_47_cast_fp16")];
+            tensor<int32, [4]> var_519_begin_0 = const()[name = string("op_519_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_519_end_0 = const()[name = string("op_519_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_519_end_mask_0 = const()[name = string("op_519_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_519_cast_fp16 = slice_by_index(begin = var_519_begin_0, end = var_519_end_0, end_mask = var_519_end_mask_0, x = obj_41_cast_fp16)[name = string("op_519_cast_fp16")];
+            tensor<int32, [4]> var_522_begin_0 = const()[name = string("op_522_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_522_end_0 = const()[name = string("op_522_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_522_end_mask_0 = const()[name = string("op_522_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_522_squeeze_mask_0 = const()[name = string("op_522_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_522_cast_fp16 = slice_by_index(begin = var_522_begin_0, end = var_522_end_0, end_mask = var_522_end_mask_0, squeeze_mask = var_522_squeeze_mask_0, x = var_519_cast_fp16)[name = string("op_522_cast_fp16")];
+            tensor<int32, [4]> var_537_begin_0 = const()[name = string("op_537_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
+            tensor<int32, [4]> var_537_end_0 = const()[name = string("op_537_end_0"), val = tensor<int32, [4]>([1, 2, 1, 1536])];
+            tensor<bool, [4]> var_537_end_mask_0 = const()[name = string("op_537_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_537_cast_fp16 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = obj_41_cast_fp16)[name = string("op_537_cast_fp16")];
+            tensor<int32, [4]> var_540_begin_0 = const()[name = string("op_540_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_540_end_0 = const()[name = string("op_540_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_540_end_mask_0 = const()[name = string("op_540_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_540_squeeze_mask_0 = const()[name = string("op_540_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_540_cast_fp16 = slice_by_index(begin = var_540_begin_0, end = var_540_end_0, end_mask = var_540_end_mask_0, squeeze_mask = var_540_squeeze_mask_0, x = var_537_cast_fp16)[name = string("op_540_cast_fp16")];
+            tensor<int32, [4]> var_555_begin_0 = const()[name = string("op_555_begin_0"), val = tensor<int32, [4]>([0, 2, 0, 0])];
+            tensor<int32, [4]> var_555_end_0 = const()[name = string("op_555_end_0"), val = tensor<int32, [4]>([1, 3, 1, 1536])];
+            tensor<bool, [4]> var_555_end_mask_0 = const()[name = string("op_555_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_555_cast_fp16 = slice_by_index(begin = var_555_begin_0, end = var_555_end_0, end_mask = var_555_end_mask_0, x = obj_41_cast_fp16)[name = string("op_555_cast_fp16")];
+            tensor<int32, [4]> var_558_begin_0 = const()[name = string("op_558_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_558_end_0 = const()[name = string("op_558_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_558_end_mask_0 = const()[name = string("op_558_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_558_squeeze_mask_0 = const()[name = string("op_558_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_558_cast_fp16 = slice_by_index(begin = var_558_begin_0, end = var_558_end_0, end_mask = var_558_end_mask_0, squeeze_mask = var_558_squeeze_mask_0, x = var_555_cast_fp16)[name = string("op_558_cast_fp16")];
+            tensor<int32, [4]> var_573_begin_0 = const()[name = string("op_573_begin_0"), val = tensor<int32, [4]>([0, 3, 0, 0])];
+            tensor<int32, [4]> var_573_end_0 = const()[name = string("op_573_end_0"), val = tensor<int32, [4]>([1, 4, 1, 1536])];
+            tensor<bool, [4]> var_573_end_mask_0 = const()[name = string("op_573_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_573_cast_fp16 = slice_by_index(begin = var_573_begin_0, end = var_573_end_0, end_mask = var_573_end_mask_0, x = obj_41_cast_fp16)[name = string("op_573_cast_fp16")];
+            tensor<int32, [4]> var_576_begin_0 = const()[name = string("op_576_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_576_end_0 = const()[name = string("op_576_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_576_end_mask_0 = const()[name = string("op_576_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_576_squeeze_mask_0 = const()[name = string("op_576_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_576_cast_fp16 = slice_by_index(begin = var_576_begin_0, end = var_576_end_0, end_mask = var_576_end_mask_0, squeeze_mask = var_576_squeeze_mask_0, x = var_573_cast_fp16)[name = string("op_576_cast_fp16")];
+            tensor<int32, [4]> var_591_begin_0 = const()[name = string("op_591_begin_0"), val = tensor<int32, [4]>([0, 4, 0, 0])];
+            tensor<int32, [4]> var_591_end_0 = const()[name = string("op_591_end_0"), val = tensor<int32, [4]>([1, 5, 1, 1536])];
+            tensor<bool, [4]> var_591_end_mask_0 = const()[name = string("op_591_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_591_cast_fp16 = slice_by_index(begin = var_591_begin_0, end = var_591_end_0, end_mask = var_591_end_mask_0, x = obj_41_cast_fp16)[name = string("op_591_cast_fp16")];
+            tensor<int32, [4]> var_594_begin_0 = const()[name = string("op_594_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_594_end_0 = const()[name = string("op_594_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_594_end_mask_0 = const()[name = string("op_594_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_594_squeeze_mask_0 = const()[name = string("op_594_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_594_cast_fp16 = slice_by_index(begin = var_594_begin_0, end = var_594_end_0, end_mask = var_594_end_mask_0, squeeze_mask = var_594_squeeze_mask_0, x = var_591_cast_fp16)[name = string("op_594_cast_fp16")];
+            tensor<int32, [4]> var_609_begin_0 = const()[name = string("op_609_begin_0"), val = tensor<int32, [4]>([0, 5, 0, 0])];
+            tensor<int32, [4]> var_609_end_0 = const()[name = string("op_609_end_0"), val = tensor<int32, [4]>([1, 6, 1, 1536])];
+            tensor<bool, [4]> var_609_end_mask_0 = const()[name = string("op_609_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_609_cast_fp16 = slice_by_index(begin = var_609_begin_0, end = var_609_end_0, end_mask = var_609_end_mask_0, x = obj_41_cast_fp16)[name = string("op_609_cast_fp16")];
+            tensor<int32, [4]> var_612_begin_0 = const()[name = string("op_612_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_612_end_0 = const()[name = string("op_612_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_612_end_mask_0 = const()[name = string("op_612_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_612_squeeze_mask_0 = const()[name = string("op_612_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_612_cast_fp16 = slice_by_index(begin = var_612_begin_0, end = var_612_end_0, end_mask = var_612_end_mask_0, squeeze_mask = var_612_squeeze_mask_0, x = var_609_cast_fp16)[name = string("op_612_cast_fp16")];
+            tensor<int32, [4]> var_627_begin_0 = const()[name = string("op_627_begin_0"), val = tensor<int32, [4]>([0, 6, 0, 0])];
+            tensor<int32, [4]> var_627_end_0 = const()[name = string("op_627_end_0"), val = tensor<int32, [4]>([1, 7, 1, 1536])];
+            tensor<bool, [4]> var_627_end_mask_0 = const()[name = string("op_627_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_627_cast_fp16 = slice_by_index(begin = var_627_begin_0, end = var_627_end_0, end_mask = var_627_end_mask_0, x = obj_41_cast_fp16)[name = string("op_627_cast_fp16")];
+            tensor<int32, [4]> var_630_begin_0 = const()[name = string("op_630_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_630_end_0 = const()[name = string("op_630_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_630_end_mask_0 = const()[name = string("op_630_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_630_squeeze_mask_0 = const()[name = string("op_630_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_630_cast_fp16 = slice_by_index(begin = var_630_begin_0, end = var_630_end_0, end_mask = var_630_end_mask_0, squeeze_mask = var_630_squeeze_mask_0, x = var_627_cast_fp16)[name = string("op_630_cast_fp16")];
+            tensor<int32, [4]> var_645_begin_0 = const()[name = string("op_645_begin_0"), val = tensor<int32, [4]>([0, 7, 0, 0])];
+            tensor<int32, [4]> var_645_end_0 = const()[name = string("op_645_end_0"), val = tensor<int32, [4]>([1, 8, 1, 1536])];
+            tensor<bool, [4]> var_645_end_mask_0 = const()[name = string("op_645_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_645_cast_fp16 = slice_by_index(begin = var_645_begin_0, end = var_645_end_0, end_mask = var_645_end_mask_0, x = obj_41_cast_fp16)[name = string("op_645_cast_fp16")];
+            tensor<int32, [4]> var_648_begin_0 = const()[name = string("op_648_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_648_end_0 = const()[name = string("op_648_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_648_end_mask_0 = const()[name = string("op_648_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_648_squeeze_mask_0 = const()[name = string("op_648_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_648_cast_fp16 = slice_by_index(begin = var_648_begin_0, end = var_648_end_0, end_mask = var_648_end_mask_0, squeeze_mask = var_648_squeeze_mask_0, x = var_645_cast_fp16)[name = string("op_648_cast_fp16")];
+            tensor<int32, [4]> var_663_begin_0 = const()[name = string("op_663_begin_0"), val = tensor<int32, [4]>([0, 8, 0, 0])];
+            tensor<int32, [4]> var_663_end_0 = const()[name = string("op_663_end_0"), val = tensor<int32, [4]>([1, 9, 1, 1536])];
+            tensor<bool, [4]> var_663_end_mask_0 = const()[name = string("op_663_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_663_cast_fp16 = slice_by_index(begin = var_663_begin_0, end = var_663_end_0, end_mask = var_663_end_mask_0, x = obj_41_cast_fp16)[name = string("op_663_cast_fp16")];
+            tensor<int32, [4]> var_666_begin_0 = const()[name = string("op_666_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_666_end_0 = const()[name = string("op_666_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_666_end_mask_0 = const()[name = string("op_666_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_666_squeeze_mask_0 = const()[name = string("op_666_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_666_cast_fp16 = slice_by_index(begin = var_666_begin_0, end = var_666_end_0, end_mask = var_666_end_mask_0, squeeze_mask = var_666_squeeze_mask_0, x = var_663_cast_fp16)[name = string("op_666_cast_fp16")];
+            tensor<int32, [4]> var_681_begin_0 = const()[name = string("op_681_begin_0"), val = tensor<int32, [4]>([0, 9, 0, 0])];
+            tensor<int32, [4]> var_681_end_0 = const()[name = string("op_681_end_0"), val = tensor<int32, [4]>([1, 10, 1, 1536])];
+            tensor<bool, [4]> var_681_end_mask_0 = const()[name = string("op_681_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_681_cast_fp16 = slice_by_index(begin = var_681_begin_0, end = var_681_end_0, end_mask = var_681_end_mask_0, x = obj_41_cast_fp16)[name = string("op_681_cast_fp16")];
+            tensor<int32, [4]> var_684_begin_0 = const()[name = string("op_684_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_684_end_0 = const()[name = string("op_684_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_684_end_mask_0 = const()[name = string("op_684_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_684_squeeze_mask_0 = const()[name = string("op_684_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_684_cast_fp16 = slice_by_index(begin = var_684_begin_0, end = var_684_end_0, end_mask = var_684_end_mask_0, squeeze_mask = var_684_squeeze_mask_0, x = var_681_cast_fp16)[name = string("op_684_cast_fp16")];
+            tensor<int32, [4]> var_699_begin_0 = const()[name = string("op_699_begin_0"), val = tensor<int32, [4]>([0, 10, 0, 0])];
+            tensor<int32, [4]> var_699_end_0 = const()[name = string("op_699_end_0"), val = tensor<int32, [4]>([1, 11, 1, 1536])];
+            tensor<bool, [4]> var_699_end_mask_0 = const()[name = string("op_699_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_699_cast_fp16 = slice_by_index(begin = var_699_begin_0, end = var_699_end_0, end_mask = var_699_end_mask_0, x = obj_41_cast_fp16)[name = string("op_699_cast_fp16")];
+            tensor<int32, [4]> var_702_begin_0 = const()[name = string("op_702_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_702_end_0 = const()[name = string("op_702_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_702_end_mask_0 = const()[name = string("op_702_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_702_squeeze_mask_0 = const()[name = string("op_702_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_702_cast_fp16 = slice_by_index(begin = var_702_begin_0, end = var_702_end_0, end_mask = var_702_end_mask_0, squeeze_mask = var_702_squeeze_mask_0, x = var_699_cast_fp16)[name = string("op_702_cast_fp16")];
+            tensor<int32, [4]> var_717_begin_0 = const()[name = string("op_717_begin_0"), val = tensor<int32, [4]>([0, 11, 0, 0])];
+            tensor<int32, [4]> var_717_end_0 = const()[name = string("op_717_end_0"), val = tensor<int32, [4]>([1, 12, 1, 1536])];
+            tensor<bool, [4]> var_717_end_mask_0 = const()[name = string("op_717_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_717_cast_fp16 = slice_by_index(begin = var_717_begin_0, end = var_717_end_0, end_mask = var_717_end_mask_0, x = obj_41_cast_fp16)[name = string("op_717_cast_fp16")];
+            tensor<int32, [4]> var_720_begin_0 = const()[name = string("op_720_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_720_end_0 = const()[name = string("op_720_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_720_end_mask_0 = const()[name = string("op_720_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_720_squeeze_mask_0 = const()[name = string("op_720_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_720_cast_fp16 = slice_by_index(begin = var_720_begin_0, end = var_720_end_0, end_mask = var_720_end_mask_0, squeeze_mask = var_720_squeeze_mask_0, x = var_717_cast_fp16)[name = string("op_720_cast_fp16")];
+            tensor<int32, [4]> var_735_begin_0 = const()[name = string("op_735_begin_0"), val = tensor<int32, [4]>([0, 12, 0, 0])];
+            tensor<int32, [4]> var_735_end_0 = const()[name = string("op_735_end_0"), val = tensor<int32, [4]>([1, 13, 1, 1536])];
+            tensor<bool, [4]> var_735_end_mask_0 = const()[name = string("op_735_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_735_cast_fp16 = slice_by_index(begin = var_735_begin_0, end = var_735_end_0, end_mask = var_735_end_mask_0, x = obj_41_cast_fp16)[name = string("op_735_cast_fp16")];
+            tensor<int32, [4]> var_738_begin_0 = const()[name = string("op_738_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_738_end_0 = const()[name = string("op_738_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_738_end_mask_0 = const()[name = string("op_738_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_738_squeeze_mask_0 = const()[name = string("op_738_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_738_cast_fp16 = slice_by_index(begin = var_738_begin_0, end = var_738_end_0, end_mask = var_738_end_mask_0, squeeze_mask = var_738_squeeze_mask_0, x = var_735_cast_fp16)[name = string("op_738_cast_fp16")];
+            tensor<int32, [4]> var_753_begin_0 = const()[name = string("op_753_begin_0"), val = tensor<int32, [4]>([0, 13, 0, 0])];
+            tensor<int32, [4]> var_753_end_0 = const()[name = string("op_753_end_0"), val = tensor<int32, [4]>([1, 14, 1, 1536])];
+            tensor<bool, [4]> var_753_end_mask_0 = const()[name = string("op_753_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_753_cast_fp16 = slice_by_index(begin = var_753_begin_0, end = var_753_end_0, end_mask = var_753_end_mask_0, x = obj_41_cast_fp16)[name = string("op_753_cast_fp16")];
+            tensor<int32, [4]> var_756_begin_0 = const()[name = string("op_756_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_756_end_0 = const()[name = string("op_756_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_756_end_mask_0 = const()[name = string("op_756_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_756_squeeze_mask_0 = const()[name = string("op_756_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_756_cast_fp16 = slice_by_index(begin = var_756_begin_0, end = var_756_end_0, end_mask = var_756_end_mask_0, squeeze_mask = var_756_squeeze_mask_0, x = var_753_cast_fp16)[name = string("op_756_cast_fp16")];
+            tensor<int32, [4]> var_771_begin_0 = const()[name = string("op_771_begin_0"), val = tensor<int32, [4]>([0, 14, 0, 0])];
+            tensor<int32, [4]> var_771_end_0 = const()[name = string("op_771_end_0"), val = tensor<int32, [4]>([1, 15, 1, 1536])];
+            tensor<bool, [4]> var_771_end_mask_0 = const()[name = string("op_771_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_771_cast_fp16 = slice_by_index(begin = var_771_begin_0, end = var_771_end_0, end_mask = var_771_end_mask_0, x = obj_41_cast_fp16)[name = string("op_771_cast_fp16")];
+            tensor<int32, [4]> var_774_begin_0 = const()[name = string("op_774_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_774_end_0 = const()[name = string("op_774_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_774_end_mask_0 = const()[name = string("op_774_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_774_squeeze_mask_0 = const()[name = string("op_774_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_774_cast_fp16 = slice_by_index(begin = var_774_begin_0, end = var_774_end_0, end_mask = var_774_end_mask_0, squeeze_mask = var_774_squeeze_mask_0, x = var_771_cast_fp16)[name = string("op_774_cast_fp16")];
+            tensor<int32, [4]> var_789_begin_0 = const()[name = string("op_789_begin_0"), val = tensor<int32, [4]>([0, 15, 0, 0])];
+            tensor<int32, [4]> var_789_end_0 = const()[name = string("op_789_end_0"), val = tensor<int32, [4]>([1, 16, 1, 1536])];
+            tensor<bool, [4]> var_789_end_mask_0 = const()[name = string("op_789_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_789_cast_fp16 = slice_by_index(begin = var_789_begin_0, end = var_789_end_0, end_mask = var_789_end_mask_0, x = obj_41_cast_fp16)[name = string("op_789_cast_fp16")];
+            tensor<int32, [4]> var_792_begin_0 = const()[name = string("op_792_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_792_end_0 = const()[name = string("op_792_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_792_end_mask_0 = const()[name = string("op_792_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_792_squeeze_mask_0 = const()[name = string("op_792_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_792_cast_fp16 = slice_by_index(begin = var_792_begin_0, end = var_792_end_0, end_mask = var_792_end_mask_0, squeeze_mask = var_792_squeeze_mask_0, x = var_789_cast_fp16)[name = string("op_792_cast_fp16")];
+            tensor<int32, [4]> var_807_begin_0 = const()[name = string("op_807_begin_0"), val = tensor<int32, [4]>([0, 16, 0, 0])];
+            tensor<int32, [4]> var_807_end_0 = const()[name = string("op_807_end_0"), val = tensor<int32, [4]>([1, 17, 1, 1536])];
+            tensor<bool, [4]> var_807_end_mask_0 = const()[name = string("op_807_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_807_cast_fp16 = slice_by_index(begin = var_807_begin_0, end = var_807_end_0, end_mask = var_807_end_mask_0, x = obj_41_cast_fp16)[name = string("op_807_cast_fp16")];
+            tensor<int32, [4]> var_810_begin_0 = const()[name = string("op_810_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_810_end_0 = const()[name = string("op_810_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_810_end_mask_0 = const()[name = string("op_810_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_810_squeeze_mask_0 = const()[name = string("op_810_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_810_cast_fp16 = slice_by_index(begin = var_810_begin_0, end = var_810_end_0, end_mask = var_810_end_mask_0, squeeze_mask = var_810_squeeze_mask_0, x = var_807_cast_fp16)[name = string("op_810_cast_fp16")];
+            tensor<int32, [4]> var_825_begin_0 = const()[name = string("op_825_begin_0"), val = tensor<int32, [4]>([0, 17, 0, 0])];
+            tensor<int32, [4]> var_825_end_0 = const()[name = string("op_825_end_0"), val = tensor<int32, [4]>([1, 18, 1, 1536])];
+            tensor<bool, [4]> var_825_end_mask_0 = const()[name = string("op_825_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_825_cast_fp16 = slice_by_index(begin = var_825_begin_0, end = var_825_end_0, end_mask = var_825_end_mask_0, x = obj_41_cast_fp16)[name = string("op_825_cast_fp16")];
+            tensor<int32, [4]> var_828_begin_0 = const()[name = string("op_828_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_828_end_0 = const()[name = string("op_828_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_828_end_mask_0 = const()[name = string("op_828_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_828_squeeze_mask_0 = const()[name = string("op_828_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_828_cast_fp16 = slice_by_index(begin = var_828_begin_0, end = var_828_end_0, end_mask = var_828_end_mask_0, squeeze_mask = var_828_squeeze_mask_0, x = var_825_cast_fp16)[name = string("op_828_cast_fp16")];
+            tensor<int32, [4]> var_843_begin_0 = const()[name = string("op_843_begin_0"), val = tensor<int32, [4]>([0, 18, 0, 0])];
+            tensor<int32, [4]> var_843_end_0 = const()[name = string("op_843_end_0"), val = tensor<int32, [4]>([1, 19, 1, 1536])];
+            tensor<bool, [4]> var_843_end_mask_0 = const()[name = string("op_843_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_843_cast_fp16 = slice_by_index(begin = var_843_begin_0, end = var_843_end_0, end_mask = var_843_end_mask_0, x = obj_41_cast_fp16)[name = string("op_843_cast_fp16")];
+            tensor<int32, [4]> var_846_begin_0 = const()[name = string("op_846_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_846_end_0 = const()[name = string("op_846_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_846_end_mask_0 = const()[name = string("op_846_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_846_squeeze_mask_0 = const()[name = string("op_846_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_846_cast_fp16 = slice_by_index(begin = var_846_begin_0, end = var_846_end_0, end_mask = var_846_end_mask_0, squeeze_mask = var_846_squeeze_mask_0, x = var_843_cast_fp16)[name = string("op_846_cast_fp16")];
+            tensor<int32, [4]> var_861_begin_0 = const()[name = string("op_861_begin_0"), val = tensor<int32, [4]>([0, 19, 0, 0])];
+            tensor<int32, [4]> var_861_end_0 = const()[name = string("op_861_end_0"), val = tensor<int32, [4]>([1, 20, 1, 1536])];
+            tensor<bool, [4]> var_861_end_mask_0 = const()[name = string("op_861_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_861_cast_fp16 = slice_by_index(begin = var_861_begin_0, end = var_861_end_0, end_mask = var_861_end_mask_0, x = obj_41_cast_fp16)[name = string("op_861_cast_fp16")];
+            tensor<int32, [4]> var_864_begin_0 = const()[name = string("op_864_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_864_end_0 = const()[name = string("op_864_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_864_end_mask_0 = const()[name = string("op_864_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_864_squeeze_mask_0 = const()[name = string("op_864_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_864_cast_fp16 = slice_by_index(begin = var_864_begin_0, end = var_864_end_0, end_mask = var_864_end_mask_0, squeeze_mask = var_864_squeeze_mask_0, x = var_861_cast_fp16)[name = string("op_864_cast_fp16")];
+            int32 var_871 = const()[name = string("op_871"), val = int32(1)];
+            bool var_872_interleave_0 = const()[name = string("op_872_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 20, 1536]> var_872_cast_fp16 = concat(axis = var_871, interleave = var_872_interleave_0, values = (var_522_cast_fp16, var_540_cast_fp16, var_558_cast_fp16, var_576_cast_fp16, var_594_cast_fp16, var_612_cast_fp16, var_630_cast_fp16, var_648_cast_fp16, var_666_cast_fp16, var_684_cast_fp16, var_702_cast_fp16, var_720_cast_fp16, var_738_cast_fp16, var_756_cast_fp16, var_774_cast_fp16, var_792_cast_fp16, var_810_cast_fp16, var_828_cast_fp16, var_846_cast_fp16, var_864_cast_fp16))[name = string("op_872_cast_fp16")];
+            bool var_875 = const()[name = string("op_875"), val = bool(false)];
+            tensor<int32, [1]> obj_axes_0 = const()[name = string("obj_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536]> alignment_heads_weights = reduce_mean(axes = obj_axes_0, keep_dims = var_875, x = var_872_cast_fp16)[name = string("obj_cast_fp16")];
+        } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights);
+}

distil-whisper_distil-large-v3/TextDecoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c1afaacaec2fac64e8867d758742347e10c849fdbf81c8761344b5c56a55b5d
+size 225873332

distil-whisper_distil-large-v3/config.json

@@ -0,0 +1 @@
+{"_name_or_path": "./distil-large-v3", "activation_dropout": 0.0, "activation_function": "gelu", "apply_spec_augment": false, "architectures": ["WhisperForConditionalGeneration"], "attention_dropout": 0.0, "begin_suppress_tokens": [220, 50257], "bos_token_id": 50257, "classifier_proj_size": 256, "d_model": 1280, "decoder_attention_heads": 20, "decoder_ffn_dim": 5120, "decoder_layerdrop": 0.0, "decoder_layers": 2, "decoder_start_token_id": 50258, "dropout": 0.0, "encoder_attention_heads": 20, "encoder_ffn_dim": 5120, "encoder_layerdrop": 0.0, "encoder_layers": 32, "eos_token_id": 50257, "init_std": 0.02, "is_encoder_decoder": true, "mask_feature_length": 10, "mask_feature_min_masks": 0, "mask_feature_prob": 0.0, "mask_time_length": 10, "mask_time_min_masks": 2, "mask_time_prob": 0.05, "max_length": 448, "max_source_positions": 1500, "max_target_positions": 448, "median_filter_width": 7, "model_type": "whisper", "num_hidden_layers": 32, "num_mel_bins": 128, "pad_token_id": 50256, "scale_embedding": false, "torch_dtype": "float16", "transformers_version": "4.38.0.dev0", "use_cache": true, "use_weighted_layer_sum": false, "vocab_size": 51866}

distil-whisper_distil-large-v3/generation_config.json

@@ -0,0 +1 @@
+{"alignment_heads": [[7, 0], [10, 17], [12, 18], [13, 12], [16, 1], [17, 14], [19, 11], [21, 4], [24, 1], [25, 6]], "begin_suppress_tokens": [220, 50257], "bos_token_id": 50257, "decoder_start_token_id": 50258, "eos_token_id": 50257, "forced_decoder_ids": [[1, null], [2, 50360]], "is_multilingual": true, "lang_to_id": {"<|af|>": 50327, "<|am|>": 50334, "<|ar|>": 50272, "<|as|>": 50350, "<|az|>": 50304, "<|ba|>": 50355, "<|be|>": 50330, "<|bg|>": 50292, "<|bn|>": 50302, "<|bo|>": 50347, "<|br|>": 50309, "<|bs|>": 50315, "<|ca|>": 50270, "<|cs|>": 50283, "<|cy|>": 50297, "<|da|>": 50285, "<|de|>": 50261, "<|el|>": 50281, "<|en|>": 50259, "<|es|>": 50262, "<|et|>": 50307, "<|eu|>": 50310, "<|fa|>": 50300, "<|fi|>": 50277, "<|fo|>": 50338, "<|fr|>": 50265, "<|gl|>": 50319, "<|gu|>": 50333, "<|haw|>": 50352, "<|ha|>": 50354, "<|he|>": 50279, "<|hi|>": 50276, "<|hr|>": 50291, "<|ht|>": 50339, "<|hu|>": 50286, "<|hy|>": 50312, "<|id|>": 50275, "<|is|>": 50311, "<|it|>": 50274, "<|ja|>": 50266, "<|jw|>": 50356, "<|ka|>": 50329, "<|kk|>": 50316, "<|km|>": 50323, "<|kn|>": 50306, "<|ko|>": 50264, "<|la|>": 50294, "<|lb|>": 50345, "<|ln|>": 50353, "<|lo|>": 50336, "<|lt|>": 50293, "<|lv|>": 50301, "<|mg|>": 50349, "<|mi|>": 50295, "<|mk|>": 50308, "<|ml|>": 50296, "<|mn|>": 50314, "<|mr|>": 50320, "<|ms|>": 50282, "<|mt|>": 50343, "<|my|>": 50346, "<|ne|>": 50313, "<|nl|>": 50271, "<|nn|>": 50342, "<|no|>": 50288, "<|oc|>": 50328, "<|pa|>": 50321, "<|pl|>": 50269, "<|ps|>": 50340, "<|pt|>": 50267, "<|ro|>": 50284, "<|ru|>": 50263, "<|sa|>": 50344, "<|sd|>": 50332, "<|si|>": 50322, "<|sk|>": 50298, "<|sl|>": 50305, "<|sn|>": 50324, "<|so|>": 50326, "<|sq|>": 50317, "<|sr|>": 50303, "<|su|>": 50357, "<|sv|>": 50273, "<|sw|>": 50318, "<|ta|>": 50287, "<|te|>": 50299, "<|tg|>": 50331, "<|th|>": 50289, "<|tk|>": 50341, "<|tl|>": 50348, "<|tr|>": 50268, "<|tt|>": 50351, "<|uk|>": 50280, "<|ur|>": 50290, "<|uz|>": 50337, "<|vi|>": 50278, "<|yi|>": 50335, "<|yo|>": 50325, "<|yue|>": 50358, "<|zh|>": 50260}, "language": "<|en|>", "max_initial_timestamp_index": 50, "max_length": 448, "no_timestamps_token_id": 50364, "pad_token_id": 50257, "prev_sot_token_id": 50362, "return_timestamps": false, "suppress_tokens": [1, 2, 7, 8, 9, 10, 14, 25, 26, 27, 28, 29, 31, 58, 59, 60, 61, 62, 63, 90, 91, 92, 93, 359, 503, 522, 542, 873, 893, 902, 918, 922, 931, 1350, 1853, 1982, 2460, 2627, 3246, 3253, 3268, 3536, 3846, 3961, 4183, 4667, 6585, 6647, 7273, 9061, 9383, 10428, 10929, 11938, 12033, 12331, 12562, 13793, 14157, 14635, 15265, 15618, 16553, 16604, 18362, 18956, 20075, 21675, 22520, 26130, 26161, 26435, 28279, 29464, 31650, 32302, 32470, 36865, 42863, 47425, 49870, 50254, 50258, 50359, 50360, 50361, 50362, 50363], "task": "transcribe", "task_to_id": {"transcribe": 50360, "translate": 50359}, "transformers_version": "4.38.0.dev0"}

distil-whisper_distil-large-v3_turbo/AudioEncoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5916646e39691156fca37ff36b96f162e80acce84cd8ee2e971115edf412a87a
+size 243

distil-whisper_distil-large-v3_turbo/AudioEncoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dbd63faaf82dd7f50bad861be88605fbdce6f59a2ced9954ef1a54a51f1e26ac
+size 434

distil-whisper_distil-large-v3_turbo/AudioEncoder.mlmodelc/metadata.json

@@ -0,0 +1,91 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 1280 × 1 × 1500)",
+        "shortDescription" : "",
+        "shape" : "[1, 1280, 1, 1500]",
+        "name" : "encoder_output_embeds",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 2 × 1280 × 1 × 1536)",
+        "shortDescription" : "",
+        "shape" : "[2, 1280, 1, 1536]",
+        "name" : "encoder_attn_key_cache",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 2 × 1280 × 1 × 1536)",
+        "shortDescription" : "",
+        "shape" : "[2, 1280, 1, 1536]",
+        "name" : "encoder_attn_value_cache",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Pad" : 2,
+      "Ios18.batchNorm" : 65,
+      "Ios18.conv" : 198,
+      "Ios18.gelu" : 34,
+      "Ios18.concat" : 674,
+      "Ios16.einsum" : 5120,
+      "Ios18.add" : 65,
+      "Ios18.softmax" : 2560,
+      "Ios18.sliceByIndex" : 4480,
+      "Ios18.layerNorm" : 65,
+      "Ios18.transpose" : 32,
+      "Ios18.mul" : 2560
+    },
+    "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.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.5.1",
+      "com.github.apple.coremltools.version" : "8.0"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 128 × 1 × 3000)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 1, 3000]",
+        "name" : "melspectrogram_features",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "AudioEncoderStateful",
+    "method" : "predict"
+  }
+]

distil-whisper_distil-large-v3_turbo/AudioEncoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b43a5d9e21e95067e0af8cf4b8fcbd16cc8e6f99993084f5e67cdf81bde16e79
+size 1287087104

distil-whisper_distil-large-v3_turbo/LICENSE_NOTICE.txt

@@ -0,0 +1,7 @@
+Argmax proprietary and confidential. Under NDA.
+
+Copyright 2024 Argmax, Inc. All rights reserved.
+
+Unauthorized access, copying, use, distribution, and or commercialization of this file, via any medium or means is strictly prohibited.
+
+Please contact Argmax for licensing information at [email protected].

distil-whisper_distil-large-v3_turbo/MelSpectrogram.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0980462db89a546e1e90888ea38e0a5ddf1f1fec84608802cdbb12f8a5cc7215
+size 243

distil-whisper_distil-large-v3_turbo/MelSpectrogram.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6475c6649047ce609e3fe84b2525843c03342820662404540baf28146c174014
+size 329

distil-whisper_distil-large-v3_turbo/MelSpectrogram.mlmodelc/metadata.json

@@ -0,0 +1,74 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 128 × 1 × 3000)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 1, 3000]",
+        "name" : "melspectrogram_features",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.mul" : 2,
+      "Ios18.square" : 2,
+      "Ios18.conv" : 2,
+      "Ios18.matmul" : 1,
+      "Ios18.expandDims" : 4,
+      "Ios18.sub" : 1,
+      "Ios18.log" : 1,
+      "Ios18.add" : 3,
+      "Ios18.sliceByIndex" : 1,
+      "Ios18.maximum" : 1,
+      "Ios18.squeeze" : 2,
+      "Ios18.reshape" : 2,
+      "Ios16.reduceMax" : 1,
+      "Identity" : 1,
+      "Pad" : 1
+    },
+    "computePrecision" : "Mixed (Float16, Float32, 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.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.5.1",
+      "com.github.apple.coremltools.version" : "8.0"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 480000)",
+        "shortDescription" : "",
+        "shape" : "[480000]",
+        "name" : "audio",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "MelSpectrogram",
+    "method" : "predict"
+  }
+]

distil-whisper_distil-large-v3_turbo/MelSpectrogram.mlmodelc/model.mil

@@ -0,0 +1,66 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3401.3.1"}, {"coremlc-version", "3401.4.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.0"}})]
+{
+    func main<ios18>(tensor<fp16, [480000]> audio) {
+            tensor<int32, [3]> var_10 = const()[name = string("op_10"), val = tensor<int32, [3]>([1, 1, 480000])];
+            tensor<fp16, [1, 1, 480000]> input_1_cast_fp16 = reshape(shape = var_10, x = audio)[name = string("input_1_cast_fp16")];
+            tensor<int32, [6]> input_3_pad_0 = const()[name = string("input_3_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 200, 200])];
+            string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("reflect")];
+            fp16 const_1_to_fp16 = const()[name = string("const_1_to_fp16"), val = fp16(0x0p+0)];
+            tensor<fp16, [1, 1, 480400]> input_3_cast_fp16 = pad(constant_val = const_1_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = string("input_3_cast_fp16")];
+            tensor<int32, [1]> var_22 = const()[name = string("op_22"), val = tensor<int32, [1]>([480400])];
+            tensor<fp16, [480400]> input_cast_fp16 = reshape(shape = var_22, x = input_3_cast_fp16)[name = string("input_cast_fp16")];
+            tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = string("expand_dims_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 480400]> expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = input_cast_fp16)[name = string("expand_dims_0_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, 480400]> expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = expand_dims_0_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, [201, 1, 400]> expand_dims_1_to_fp16 = const()[name = string("expand_dims_1_to_fp16"), val = tensor<fp16, [201, 1, 400]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 201, 3001]> 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, [201, 1, 400]> expand_dims_2_to_fp16 = const()[name = string("expand_dims_2_to_fp16"), val = tensor<fp16, [201, 1, 400]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160960)))];
+            tensor<fp16, [1, 201, 3001]> 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")];
+            tensor<int32, [1]> squeeze_0_axes_0 = const()[name = string("squeeze_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [201, 3001]> squeeze_0_cast_fp16 = squeeze(axes = squeeze_0_axes_0, x = conv_0_cast_fp16)[name = string("squeeze_0_cast_fp16")];
+            tensor<int32, [1]> squeeze_1_axes_0 = const()[name = string("squeeze_1_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [201, 3001]> squeeze_1_cast_fp16 = squeeze(axes = squeeze_1_axes_0, x = conv_1_cast_fp16)[name = string("squeeze_1_cast_fp16")];
+            tensor<fp16, [201, 3001]> square_0_cast_fp16 = square(x = squeeze_0_cast_fp16)[name = string("square_0_cast_fp16")];
+            tensor<fp16, [201, 3001]> square_1_cast_fp16 = square(x = squeeze_1_cast_fp16)[name = string("square_1_cast_fp16")];
+            tensor<fp16, [201, 3001]> add_1_cast_fp16 = add(x = square_0_cast_fp16, y = square_1_cast_fp16)[name = string("add_1_cast_fp16")];
+            tensor<fp16, [201, 3001]> magnitudes_1_cast_fp16 = identity(x = add_1_cast_fp16)[name = string("magnitudes_1_cast_fp16")];
+            tensor<int32, [2]> magnitudes_begin_0 = const()[name = string("magnitudes_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> magnitudes_end_0 = const()[name = string("magnitudes_end_0"), val = tensor<int32, [2]>([201, 3000])];
+            tensor<bool, [2]> magnitudes_end_mask_0 = const()[name = string("magnitudes_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp16, [201, 3000]> magnitudes_cast_fp16 = slice_by_index(begin = magnitudes_begin_0, end = magnitudes_end_0, end_mask = magnitudes_end_mask_0, x = magnitudes_1_cast_fp16)[name = string("magnitudes_cast_fp16")];
+            bool mel_spec_1_transpose_x_0 = const()[name = string("mel_spec_1_transpose_x_0"), val = bool(false)];
+            bool mel_spec_1_transpose_y_0 = const()[name = string("mel_spec_1_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [128, 201]> mel_filters_to_fp16 = const()[name = string("mel_filters_to_fp16"), val = tensor<fp16, [128, 201]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321856)))];
+            tensor<fp16, [128, 3000]> mel_spec_1_cast_fp16 = matmul(transpose_x = mel_spec_1_transpose_x_0, transpose_y = mel_spec_1_transpose_y_0, x = mel_filters_to_fp16, y = magnitudes_cast_fp16)[name = string("mel_spec_1_cast_fp16")];
+            fp16 var_41_to_fp16 = const()[name = string("op_41_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [128, 3000]> mel_spec_cast_fp16 = add(x = mel_spec_1_cast_fp16, y = var_41_to_fp16)[name = string("mel_spec_cast_fp16")];
+            fp32 log_0_epsilon_0 = const()[name = string("log_0_epsilon_0"), val = fp32(0x1p-149)];
+            tensor<fp16, [128, 3000]> log_0_cast_fp16 = log(epsilon = log_0_epsilon_0, x = mel_spec_cast_fp16)[name = string("log_0_cast_fp16")];
+            fp16 mul_0_y_0_to_fp16 = const()[name = string("mul_0_y_0_to_fp16"), val = fp16(0x1.bccp-2)];
+            tensor<fp16, [128, 3000]> mul_0_cast_fp16 = mul(x = log_0_cast_fp16, y = mul_0_y_0_to_fp16)[name = string("mul_0_cast_fp16")];
+            bool var_44_keep_dims_0 = const()[name = string("op_44_keep_dims_0"), val = bool(false)];
+            fp16 var_44_cast_fp16 = reduce_max(keep_dims = var_44_keep_dims_0, x = mul_0_cast_fp16)[name = string("op_44_cast_fp16")];
+            fp16 var_46_to_fp16 = const()[name = string("op_46_to_fp16"), val = fp16(0x1p+3)];
+            fp16 var_47_cast_fp16 = sub(x = var_44_cast_fp16, y = var_46_to_fp16)[name = string("op_47_cast_fp16")];
+            tensor<fp16, [128, 3000]> log_spec_3_cast_fp16 = maximum(x = mul_0_cast_fp16, y = var_47_cast_fp16)[name = string("log_spec_3_cast_fp16")];
+            fp16 var_50_to_fp16 = const()[name = string("op_50_to_fp16"), val = fp16(0x1p+2)];
+            tensor<fp16, [128, 3000]> var_51_cast_fp16 = add(x = log_spec_3_cast_fp16, y = var_50_to_fp16)[name = string("op_51_cast_fp16")];
+            fp16 _inversed_log_spec_y_0_to_fp16 = const()[name = string("_inversed_log_spec_y_0_to_fp16"), val = fp16(0x1p-2)];
+            tensor<fp16, [128, 3000]> _inversed_log_spec_cast_fp16 = mul(x = var_51_cast_fp16, y = _inversed_log_spec_y_0_to_fp16)[name = string("_inversed_log_spec_cast_fp16")];
+            tensor<int32, [1]> var_55_axes_0 = const()[name = string("op_55_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 128, 3000]> var_55_cast_fp16 = expand_dims(axes = var_55_axes_0, x = _inversed_log_spec_cast_fp16)[name = string("op_55_cast_fp16")];
+            tensor<int32, [1]> var_62_axes_0 = const()[name = string("op_62_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 128, 1, 3000]> melspectrogram_features = expand_dims(axes = var_62_axes_0, x = var_55_cast_fp16)[name = string("op_62_cast_fp16")];
+        } -> (melspectrogram_features);
+}

distil-whisper_distil-large-v3_turbo/MelSpectrogram.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:009d9fb8f6b589accfa08cebf1c712ef07c3405229ce3cfb3a57ee033c9d8a49
+size 373376

distil-whisper_distil-large-v3_turbo/TextDecoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77cb1b565a336e7fc01586698e50aa32d9a2a8f1ca5c439172564f4af0515f5d
+size 243

distil-whisper_distil-large-v3_turbo/TextDecoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a5e6f62b5ae897c8f846e22cacbe7d4f7d6bdbeb5f46366e2387f1082676b62
+size 754

distil-whisper_distil-large-v3_turbo/TextDecoder.mlmodelc/metadata.json

@@ -0,0 +1,183 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 1 × 51866)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 51866]",
+        "name" : "logits",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 2560 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 2560, 1, 1]",
+        "name" : "key_cache_updates",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 2560 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 2560, 1, 1]",
+        "name" : "value_cache_updates",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 1536)",
+        "shortDescription" : "",
+        "shape" : "[1, 1536]",
+        "name" : "alignment_heads_weights",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.expandDims" : 8,
+      "Ios18.softmax" : 4,
+      "Ios18.mul" : 8,
+      "Ios18.matmul" : 8,
+      "Ios18.batchNorm" : 7,
+      "Ios16.reduceMean" : 1,
+      "Split" : 2,
+      "Ios18.readState" : 5,
+      "Ios18.gather" : 2,
+      "Ios18.add" : 15,
+      "Ios18.layerNorm" : 7,
+      "Ios18.reshape" : 16,
+      "Ios18.linear" : 1,
+      "Ios18.conv" : 16,
+      "Ios18.gelu" : 2,
+      "Ios18.concat" : 3,
+      "Ios18.cast" : 1,
+      "Ios18.transpose" : 1,
+      "Ios18.sliceByIndex" : 44,
+      "Ios18.squeeze" : 1
+    },
+    "computePrecision" : "Mixed (Float16, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 1 × 1536)",
+        "shortDescription" : "",
+        "shape" : "[1, 1536]",
+        "name" : "encoder_attn_key_padding_mask",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 2 × 1280 × 1 × 1536)",
+        "shortDescription" : "",
+        "shape" : "[2, 1280, 1, 1536]",
+        "name" : "encoder_attn_key_cache",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 2 × 1280 × 1 × 1536)",
+        "shortDescription" : "",
+        "shape" : "[2, 1280, 1, 1536]",
+        "name" : "encoder_attn_value_cache",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 2 × 1280 × 1 × 448)",
+        "shortDescription" : "",
+        "shape" : "[2, 1280, 1, 448]",
+        "name" : "self_attn_key_cache",
+        "type" : "State"
+      },
+      {
+        "dataType" : "Float16",
+        "isOptional" : "0",
+        "formattedType" : "State (Float16 2 × 1280 × 1 × 448)",
+        "shortDescription" : "",
+        "shape" : "[2, 1280, 1, 448]",
+        "name" : "self_attn_value_cache",
+        "type" : "State"
+      }
+    ],
+    "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.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.5.1",
+      "com.github.apple.coremltools.version" : "8.0"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "input_ids",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "cache_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 448)",
+        "shortDescription" : "",
+        "shape" : "[1, 448]",
+        "name" : "kv_cache_update_mask",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 448)",
+        "shortDescription" : "",
+        "shape" : "[1, 448]",
+        "name" : "decoder_key_padding_mask",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "TextDecoderStateful",
+    "method" : "predict"
+  }
+]

distil-whisper_distil-large-v3_turbo/TextDecoder.mlmodelc/model.mil

@@ -0,0 +1,529 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3401.3.1"}, {"coremlc-version", "3401.4.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.0"}})]
+{
+    func main<ios18>(tensor<int32, [1]> cache_length, tensor<fp16, [1, 448]> decoder_key_padding_mask, state<tensor<fp16, [2, 1280, 1, 1536]>> encoder_attn_key_cache, state<tensor<fp16, [1, 1536]>> encoder_attn_key_padding_mask, state<tensor<fp16, [2, 1280, 1, 1536]>> encoder_attn_value_cache, tensor<int32, [1]> input_ids, tensor<fp16, [1, 448]> kv_cache_update_mask, state<tensor<fp16, [2, 1280, 1, 448]>> self_attn_key_cache, state<tensor<fp16, [2, 1280, 1, 448]>> self_attn_value_cache) {
+            int32 var_22_axis_0 = const()[name = string("op_22_axis_0"), val = int32(0)];
+            int32 var_22_batch_dims_0 = const()[name = string("op_22_batch_dims_0"), val = int32(0)];
+            bool var_22_validate_indices_0 = const()[name = string("op_22_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [51866, 1280]> embed_tokens_weight_to_fp16 = const()[name = string("embed_tokens_weight_to_fp16"), val = tensor<fp16, [51866, 1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            tensor<fp16, [1, 1280]> var_22_cast_fp16 = gather(axis = var_22_axis_0, batch_dims = var_22_batch_dims_0, indices = input_ids, validate_indices = var_22_validate_indices_0, x = embed_tokens_weight_to_fp16)[name = string("op_22_cast_fp16")];
+            int32 var_26_axis_0 = const()[name = string("op_26_axis_0"), val = int32(0)];
+            int32 var_26_batch_dims_0 = const()[name = string("op_26_batch_dims_0"), val = int32(0)];
+            bool var_26_validate_indices_0 = const()[name = string("op_26_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [448, 1280]> embed_positions_weight_to_fp16 = const()[name = string("embed_positions_weight_to_fp16"), val = tensor<fp16, [448, 1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132777088)))];
+            string cache_length_to_uint16_dtype_0 = const()[name = string("cache_length_to_uint16_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1]> cache_length_to_uint16 = cast(dtype = cache_length_to_uint16_dtype_0, x = cache_length)[name = string("cast_43")];
+            tensor<fp16, [1, 1280]> var_26_cast_fp16_cast_uint16 = gather(axis = var_26_axis_0, batch_dims = var_26_batch_dims_0, indices = cache_length_to_uint16, validate_indices = var_26_validate_indices_0, x = embed_positions_weight_to_fp16)[name = string("op_26_cast_fp16_cast_uint16")];
+            tensor<fp16, [1, 1280]> hidden_states_1_cast_fp16 = add(x = var_22_cast_fp16, y = var_26_cast_fp16_cast_uint16)[name = string("hidden_states_1_cast_fp16")];
+            tensor<int32, [1]> var_40_axes_0 = const()[name = string("op_40_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1280, 1]> var_40_cast_fp16 = expand_dims(axes = var_40_axes_0, x = hidden_states_1_cast_fp16)[name = string("op_40_cast_fp16")];
+            tensor<int32, [1]> inputs_1_axes_0 = const()[name = string("inputs_1_axes_0"), val = tensor<int32, [1]>([3])];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_40_cast_fp16)[name = string("inputs_1_cast_fp16")];
+            tensor<fp16, [2, 1280, 1, 448]> read_state_0 = read_state(input = self_attn_key_cache)[name = string("read_state_0")];
+            tensor<int32, [2]> tile_0 = const()[name = string("tile_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_45_axis_0 = const()[name = string("op_45_axis_0"), val = int32(0)];
+            tensor<fp16, [1, 1280, 1, 448]> var_45_cast_fp16_0, tensor<fp16, [1, 1280, 1, 448]> var_45_cast_fp16_1 = split(axis = var_45_axis_0, split_sizes = tile_0, x = read_state_0)[name = string("op_45_cast_fp16")];
+            tensor<fp16, [2, 1280, 1, 448]> read_state_1 = read_state(input = self_attn_value_cache)[name = string("read_state_1")];
+            tensor<int32, [2]> tile_1 = const()[name = string("tile_1"), val = tensor<int32, [2]>([1, 1])];
+            int32 var_50_axis_0 = const()[name = string("op_50_axis_0"), val = int32(0)];
+            tensor<fp16, [1, 1280, 1, 448]> var_50_cast_fp16_0, tensor<fp16, [1, 1280, 1, 448]> var_50_cast_fp16_1 = split(axis = var_50_axis_0, split_sizes = tile_1, x = read_state_1)[name = string("op_50_cast_fp16")];
+            tensor<fp16, [2, 1280, 1, 1536]> read_state_2 = read_state(input = encoder_attn_key_cache)[name = string("read_state_2")];
+            tensor<int32, [4]> obj_17_begin_0 = const()[name = string("obj_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> obj_17_end_0 = const()[name = string("obj_17_end_0"), val = tensor<int32, [4]>([1, 1280, 1, 1536])];
+            tensor<bool, [4]> obj_17_end_mask_0 = const()[name = string("obj_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 1280, 1, 1536]> obj_17_cast_fp16 = slice_by_index(begin = obj_17_begin_0, end = obj_17_end_0, end_mask = obj_17_end_mask_0, x = read_state_2)[name = string("obj_17_cast_fp16")];
+            tensor<fp16, [2, 1280, 1, 1536]> read_state_3 = read_state(input = encoder_attn_value_cache)[name = string("read_state_3")];
+            tensor<int32, [4]> obj_19_begin_0 = const()[name = string("obj_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> obj_19_end_0 = const()[name = string("obj_19_end_0"), val = tensor<int32, [4]>([1, 1280, 1, 1536])];
+            tensor<bool, [4]> obj_19_end_mask_0 = const()[name = string("obj_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 1280, 1, 1536]> obj_19_cast_fp16 = slice_by_index(begin = obj_19_begin_0, end = obj_19_end_0, end_mask = obj_19_end_mask_0, x = read_state_3)[name = string("obj_19_cast_fp16")];
+            int32 var_68 = const()[name = string("op_68"), val = int32(3)];
+            tensor<int32, [1]> out_1_axes_0 = const()[name = string("out_1_axes_0"), val = tensor<int32, [1]>([1])];
+            fp16 var_93_to_fp16 = const()[name = string("op_93_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_93_to_fp16, x = inputs_1_cast_fp16)[name = string("out_1_cast_fp16")];
+            tensor<fp16, [1280]> obj_5_mean_0_to_fp16 = const()[name = string("obj_5_mean_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133924032)))];
+            tensor<fp16, [1280]> obj_5_variance_0_to_fp16 = const()[name = string("obj_5_variance_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133926656)))];
+            tensor<fp16, [1280]> obj_5_gamma_0_to_fp16 = const()[name = string("obj_5_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133929280)))];
+            tensor<fp16, [1280]> obj_5_beta_0_to_fp16 = const()[name = string("obj_5_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133931904)))];
+            fp16 obj_5_epsilon_0_to_fp16 = const()[name = string("obj_5_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> obj_5_cast_fp16 = batch_norm(beta = obj_5_beta_0_to_fp16, epsilon = obj_5_epsilon_0_to_fp16, gamma = obj_5_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_1_cast_fp16)[name = string("obj_5_cast_fp16")];
+            string query_1_pad_type_0 = const()[name = string("query_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> query_1_strides_0 = const()[name = string("query_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> query_1_pad_0 = const()[name = string("query_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> query_1_dilations_0 = const()[name = string("query_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 query_1_groups_0 = const()[name = string("query_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133934528)))];
+            tensor<fp16, [1280]> layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137211392)))];
+            tensor<fp16, [1, 1280, 1, 1]> query_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_bias_to_fp16, dilations = query_1_dilations_0, groups = query_1_groups_0, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = query_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = string("query_1_cast_fp16")];
+            string current_key_1_pad_type_0 = const()[name = string("current_key_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> current_key_1_strides_0 = const()[name = string("current_key_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> current_key_1_pad_0 = const()[name = string("current_key_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> current_key_1_dilations_0 = const()[name = string("current_key_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 current_key_1_groups_0 = const()[name = string("current_key_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137214016)))];
+            tensor<fp16, [1, 1280, 1, 1]> current_key_1_cast_fp16 = conv(dilations = current_key_1_dilations_0, groups = current_key_1_groups_0, pad = current_key_1_pad_0, pad_type = current_key_1_pad_type_0, strides = current_key_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = string("current_key_1_cast_fp16")];
+            string current_value_1_pad_type_0 = const()[name = string("current_value_1_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> current_value_1_strides_0 = const()[name = string("current_value_1_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> current_value_1_pad_0 = const()[name = string("current_value_1_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> current_value_1_dilations_0 = const()[name = string("current_value_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 current_value_1_groups_0 = const()[name = string("current_value_1_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140490880)))];
+            tensor<fp16, [1280]> layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143767744)))];
+            tensor<fp16, [1, 1280, 1, 1]> current_value_1_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_bias_to_fp16, dilations = current_value_1_dilations_0, groups = current_value_1_groups_0, pad = current_value_1_pad_0, pad_type = current_value_1_pad_type_0, strides = current_value_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = obj_5_cast_fp16)[name = string("current_value_1_cast_fp16")];
+            tensor<int32, [1]> var_128_axes_0 = const()[name = string("op_128_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 448]> var_128_cast_fp16 = expand_dims(axes = var_128_axes_0, x = kv_cache_update_mask)[name = string("op_128_cast_fp16")];
+            tensor<int32, [1]> var_129_axes_0 = const()[name = string("op_129_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 448]> var_129_cast_fp16 = expand_dims(axes = var_129_axes_0, x = var_128_cast_fp16)[name = string("op_129_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 448]> var_131_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_129_cast_fp16)[name = string("op_131_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 448]> key_1_cast_fp16 = add(x = var_45_cast_fp16_0, y = var_131_cast_fp16)[name = string("key_1_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 448]> var_133_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_129_cast_fp16)[name = string("op_133_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 448]> value_1_cast_fp16 = add(x = var_50_cast_fp16_0, y = var_133_cast_fp16)[name = string("value_1_cast_fp16")];
+            tensor<int32, [4]> var_136 = const()[name = string("op_136"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1]> mh_q_1_cast_fp16 = reshape(shape = var_136, x = query_1_cast_fp16)[name = string("mh_q_1_cast_fp16")];
+            fp16 var_138_to_fp16 = const()[name = string("op_138_to_fp16"), val = fp16(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1]> var_139_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_138_to_fp16)[name = string("op_139_cast_fp16")];
+            tensor<int32, [4]> var_140 = const()[name = string("op_140"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 448]> var_141_cast_fp16 = reshape(shape = var_140, x = key_1_cast_fp16)[name = string("op_141_cast_fp16")];
+            bool mh_w_1_transpose_x_0 = const()[name = string("mh_w_1_transpose_x_0"), val = bool(true)];
+            bool mh_w_1_transpose_y_0 = const()[name = string("mh_w_1_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 20, 1, 448]> mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_139_cast_fp16, y = var_141_cast_fp16)[name = string("mh_w_1_cast_fp16")];
+            tensor<int32, [1]> var_145_axes_0 = const()[name = string("op_145_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 448]> var_145_cast_fp16 = expand_dims(axes = var_145_axes_0, x = decoder_key_padding_mask)[name = string("op_145_cast_fp16")];
+            tensor<int32, [1]> var_146_axes_0 = const()[name = string("op_146_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 448]> var_146_cast_fp16 = expand_dims(axes = var_146_axes_0, x = var_145_cast_fp16)[name = string("op_146_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 448]> mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_146_cast_fp16)[name = string("mh_w_3_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 448]> var_149_cast_fp16 = softmax(axis = var_68, x = mh_w_3_cast_fp16)[name = string("op_149_cast_fp16")];
+            tensor<int32, [4]> var_150 = const()[name = string("op_150"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 448]> var_151_cast_fp16 = reshape(shape = var_150, x = value_1_cast_fp16)[name = string("op_151_cast_fp16")];
+            bool attn_1_transpose_x_0 = const()[name = string("attn_1_transpose_x_0"), val = bool(false)];
+            bool attn_1_transpose_y_0 = const()[name = string("attn_1_transpose_y_0"), val = bool(true)];
+            tensor<fp16, [1, 20, 64, 1]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_151_cast_fp16, y = var_149_cast_fp16)[name = string("attn_1_cast_fp16")];
+            tensor<int32, [4]> var_154 = const()[name = string("op_154"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1]> input_1_cast_fp16 = reshape(shape = var_154, x = attn_1_cast_fp16)[name = string("input_1_cast_fp16")];
+            string obj_11_pad_type_0 = const()[name = string("obj_11_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> obj_11_strides_0 = const()[name = string("obj_11_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> obj_11_pad_0 = const()[name = string("obj_11_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> obj_11_dilations_0 = const()[name = string("obj_11_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 obj_11_groups_0 = const()[name = string("obj_11_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143770368)))];
+            tensor<fp16, [1280]> layers_0_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147047232)))];
+            tensor<fp16, [1, 1280, 1, 1]> obj_11_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_bias_to_fp16, dilations = obj_11_dilations_0, groups = obj_11_groups_0, pad = obj_11_pad_0, pad_type = obj_11_pad_type_0, strides = obj_11_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("obj_11_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_11_cast_fp16)[name = string("inputs_3_cast_fp16")];
+            tensor<int32, [1]> out_3_axes_0 = const()[name = string("out_3_axes_0"), val = tensor<int32, [1]>([1])];
+            fp16 var_176_to_fp16 = const()[name = string("op_176_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_176_to_fp16, x = inputs_3_cast_fp16)[name = string("out_3_cast_fp16")];
+            tensor<fp16, [1280]> obj_13_gamma_0_to_fp16 = const()[name = string("obj_13_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147049856)))];
+            tensor<fp16, [1280]> obj_13_beta_0_to_fp16 = const()[name = string("obj_13_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147052480)))];
+            fp16 obj_13_epsilon_0_to_fp16 = const()[name = string("obj_13_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> obj_13_cast_fp16 = batch_norm(beta = obj_13_beta_0_to_fp16, epsilon = obj_13_epsilon_0_to_fp16, gamma = obj_13_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_3_cast_fp16)[name = string("obj_13_cast_fp16")];
+            string query_3_pad_type_0 = const()[name = string("query_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> query_3_strides_0 = const()[name = string("query_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> query_3_pad_0 = const()[name = string("query_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> query_3_dilations_0 = const()[name = string("query_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 query_3_groups_0 = const()[name = string("query_3_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147055104)))];
+            tensor<fp16, [1280]> layers_0_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("layers_0_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150331968)))];
+            tensor<fp16, [1, 1280, 1, 1]> query_3_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_bias_to_fp16, dilations = query_3_dilations_0, groups = query_3_groups_0, pad = query_3_pad_0, pad_type = query_3_pad_type_0, strides = query_3_strides_0, weight = layers_0_encoder_attn_q_proj_weight_to_fp16, x = obj_13_cast_fp16)[name = string("query_3_cast_fp16")];
+            tensor<int32, [4]> var_196 = const()[name = string("op_196"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1]> mh_q_3_cast_fp16 = reshape(shape = var_196, x = query_3_cast_fp16)[name = string("mh_q_3_cast_fp16")];
+            fp16 var_198_to_fp16 = const()[name = string("op_198_to_fp16"), val = fp16(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1]> var_199_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_198_to_fp16)[name = string("op_199_cast_fp16")];
+            tensor<int32, [4]> var_200 = const()[name = string("op_200"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1536]> var_201_cast_fp16 = reshape(shape = var_200, x = obj_17_cast_fp16)[name = string("op_201_cast_fp16")];
+            bool mh_w_5_transpose_x_0 = const()[name = string("mh_w_5_transpose_x_0"), val = bool(true)];
+            bool mh_w_5_transpose_y_0 = const()[name = string("mh_w_5_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 20, 1, 1536]> mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_199_cast_fp16, y = var_201_cast_fp16)[name = string("mh_w_5_cast_fp16")];
+            tensor<fp16, [1, 1536]> read_state_4 = read_state(input = encoder_attn_key_padding_mask)[name = string("read_state_4")];
+            tensor<int32, [1]> var_205_axes_0 = const()[name = string("op_205_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1536]> var_205_cast_fp16 = expand_dims(axes = var_205_axes_0, x = read_state_4)[name = string("op_205_cast_fp16")];
+            tensor<int32, [1]> var_206_axes_0 = const()[name = string("op_206_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1, 1, 1536]> var_206_cast_fp16 = expand_dims(axes = var_206_axes_0, x = var_205_cast_fp16)[name = string("op_206_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 1536]> mh_w_7_cast_fp16 = add(x = mh_w_5_cast_fp16, y = var_206_cast_fp16)[name = string("mh_w_7_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 1536]> obj_23_cast_fp16 = softmax(axis = var_68, x = mh_w_7_cast_fp16)[name = string("obj_23_cast_fp16")];
+            tensor<int32, [4]> var_210 = const()[name = string("op_210"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1536]> var_211_cast_fp16 = reshape(shape = var_210, x = obj_19_cast_fp16)[name = string("op_211_cast_fp16")];
+            bool attn_3_transpose_x_0 = const()[name = string("attn_3_transpose_x_0"), val = bool(false)];
+            bool attn_3_transpose_y_0 = const()[name = string("attn_3_transpose_y_0"), val = bool(true)];
+            tensor<fp16, [1, 20, 64, 1]> attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_211_cast_fp16, y = obj_23_cast_fp16)[name = string("attn_3_cast_fp16")];
+            tensor<int32, [4]> var_214 = const()[name = string("op_214"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1]> input_3_cast_fp16 = reshape(shape = var_214, x = attn_3_cast_fp16)[name = string("input_3_cast_fp16")];
+            string obj_21_pad_type_0 = const()[name = string("obj_21_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> obj_21_strides_0 = const()[name = string("obj_21_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> obj_21_pad_0 = const()[name = string("obj_21_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> obj_21_dilations_0 = const()[name = string("obj_21_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 obj_21_groups_0 = const()[name = string("obj_21_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_0_encoder_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(150334592)))];
+            tensor<fp16, [1280]> layers_0_encoder_attn_o_proj_bias_to_fp16 = const()[name = string("layers_0_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153611456)))];
+            tensor<fp16, [1, 1280, 1, 1]> obj_21_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_bias_to_fp16, dilations = obj_21_dilations_0, groups = obj_21_groups_0, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = obj_21_strides_0, weight = layers_0_encoder_attn_o_proj_weight_to_fp16, x = input_3_cast_fp16)[name = string("obj_21_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_21_cast_fp16)[name = string("inputs_5_cast_fp16")];
+            tensor<int32, [1]> out_5_axes_0 = const()[name = string("out_5_axes_0"), val = tensor<int32, [1]>([1])];
+            fp16 var_232_to_fp16 = const()[name = string("op_232_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_232_to_fp16, x = inputs_5_cast_fp16)[name = string("out_5_cast_fp16")];
+            tensor<fp16, [1280]> input_5_gamma_0_to_fp16 = const()[name = string("input_5_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153614080)))];
+            tensor<fp16, [1280]> input_5_beta_0_to_fp16 = const()[name = string("input_5_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153616704)))];
+            fp16 input_5_epsilon_0_to_fp16 = const()[name = string("input_5_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> input_5_cast_fp16 = batch_norm(beta = input_5_beta_0_to_fp16, epsilon = input_5_epsilon_0_to_fp16, gamma = input_5_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_5_cast_fp16)[name = string("input_5_cast_fp16")];
+            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(1)];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_0_fc1_weight_to_fp16 = const()[name = string("layers_0_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153619328)))];
+            tensor<fp16, [5120]> layers_0_fc1_bias_to_fp16 = const()[name = string("layers_0_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166726592)))];
+            tensor<fp16, [1, 5120, 1, 1]> input_7_cast_fp16 = conv(bias = layers_0_fc1_bias_to_fp16, dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = layers_0_fc1_weight_to_fp16, x = input_5_cast_fp16)[name = string("input_7_cast_fp16")];
+            string input_9_mode_0 = const()[name = string("input_9_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1]> input_9_cast_fp16 = gelu(mode = input_9_mode_0, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")];
+            string hidden_states_3_pad_type_0 = const()[name = string("hidden_states_3_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> hidden_states_3_strides_0 = const()[name = string("hidden_states_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> hidden_states_3_pad_0 = const()[name = string("hidden_states_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> hidden_states_3_dilations_0 = const()[name = string("hidden_states_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 hidden_states_3_groups_0 = const()[name = string("hidden_states_3_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_0_fc2_weight_to_fp16 = const()[name = string("layers_0_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(166736896)))];
+            tensor<fp16, [1280]> layers_0_fc2_bias_to_fp16 = const()[name = string("layers_0_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179844160)))];
+            tensor<fp16, [1, 1280, 1, 1]> hidden_states_3_cast_fp16 = conv(bias = layers_0_fc2_bias_to_fp16, dilations = hidden_states_3_dilations_0, groups = hidden_states_3_groups_0, pad = hidden_states_3_pad_0, pad_type = hidden_states_3_pad_type_0, strides = hidden_states_3_strides_0, weight = layers_0_fc2_weight_to_fp16, x = input_9_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = string("inputs_7_cast_fp16")];
+            tensor<int32, [4]> obj_35_begin_0 = const()[name = string("obj_35_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> obj_35_end_0 = const()[name = string("obj_35_end_0"), val = tensor<int32, [4]>([2, 1280, 1, 1536])];
+            tensor<bool, [4]> obj_35_end_mask_0 = const()[name = string("obj_35_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 1280, 1, 1536]> obj_35_cast_fp16 = slice_by_index(begin = obj_35_begin_0, end = obj_35_end_0, end_mask = obj_35_end_mask_0, x = read_state_2)[name = string("obj_35_cast_fp16")];
+            tensor<int32, [4]> obj_37_begin_0 = const()[name = string("obj_37_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> obj_37_end_0 = const()[name = string("obj_37_end_0"), val = tensor<int32, [4]>([2, 1280, 1, 1536])];
+            tensor<bool, [4]> obj_37_end_mask_0 = const()[name = string("obj_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<fp16, [1, 1280, 1, 1536]> obj_37_cast_fp16 = slice_by_index(begin = obj_37_begin_0, end = obj_37_end_0, end_mask = obj_37_end_mask_0, x = read_state_3)[name = string("obj_37_cast_fp16")];
+            int32 var_277 = const()[name = string("op_277"), val = int32(3)];
+            tensor<int32, [1]> out_7_axes_0 = const()[name = string("out_7_axes_0"), val = tensor<int32, [1]>([1])];
+            fp16 var_302_to_fp16 = const()[name = string("op_302_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_302_to_fp16, x = inputs_7_cast_fp16)[name = string("out_7_cast_fp16")];
+            tensor<fp16, [1280]> obj_25_gamma_0_to_fp16 = const()[name = string("obj_25_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179846784)))];
+            tensor<fp16, [1280]> obj_25_beta_0_to_fp16 = const()[name = string("obj_25_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179849408)))];
+            fp16 obj_25_epsilon_0_to_fp16 = const()[name = string("obj_25_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> obj_25_cast_fp16 = batch_norm(beta = obj_25_beta_0_to_fp16, epsilon = obj_25_epsilon_0_to_fp16, gamma = obj_25_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_7_cast_fp16)[name = string("obj_25_cast_fp16")];
+            string query_5_pad_type_0 = const()[name = string("query_5_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> query_5_strides_0 = const()[name = string("query_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> query_5_pad_0 = const()[name = string("query_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> query_5_dilations_0 = const()[name = string("query_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 query_5_groups_0 = const()[name = string("query_5_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(179852032)))];
+            tensor<fp16, [1280]> layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183128896)))];
+            tensor<fp16, [1, 1280, 1, 1]> query_5_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_bias_to_fp16, dilations = query_5_dilations_0, groups = query_5_groups_0, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = query_5_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = string("query_5_cast_fp16")];
+            string current_key_pad_type_0 = const()[name = string("current_key_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> current_key_strides_0 = const()[name = string("current_key_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> current_key_pad_0 = const()[name = string("current_key_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> current_key_dilations_0 = const()[name = string("current_key_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 current_key_groups_0 = const()[name = string("current_key_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183131520)))];
+            tensor<fp16, [1, 1280, 1, 1]> current_key_cast_fp16 = conv(dilations = current_key_dilations_0, groups = current_key_groups_0, pad = current_key_pad_0, pad_type = current_key_pad_type_0, strides = current_key_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = string("current_key_cast_fp16")];
+            string current_value_pad_type_0 = const()[name = string("current_value_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> current_value_strides_0 = const()[name = string("current_value_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> current_value_pad_0 = const()[name = string("current_value_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> current_value_dilations_0 = const()[name = string("current_value_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 current_value_groups_0 = const()[name = string("current_value_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186408384)))];
+            tensor<fp16, [1280]> layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = string("layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189685248)))];
+            tensor<fp16, [1, 1280, 1, 1]> current_value_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_bias_to_fp16, dilations = current_value_dilations_0, groups = current_value_groups_0, pad = current_value_pad_0, pad_type = current_value_pad_type_0, strides = current_value_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = obj_25_cast_fp16)[name = string("current_value_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 448]> var_340_cast_fp16 = mul(x = current_key_cast_fp16, y = var_129_cast_fp16)[name = string("op_340_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 448]> key_cast_fp16 = add(x = var_45_cast_fp16_1, y = var_340_cast_fp16)[name = string("key_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 448]> var_342_cast_fp16 = mul(x = current_value_cast_fp16, y = var_129_cast_fp16)[name = string("op_342_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 448]> value_cast_fp16 = add(x = var_50_cast_fp16_1, y = var_342_cast_fp16)[name = string("value_cast_fp16")];
+            tensor<int32, [4]> var_345 = const()[name = string("op_345"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1]> mh_q_5_cast_fp16 = reshape(shape = var_345, x = query_5_cast_fp16)[name = string("mh_q_5_cast_fp16")];
+            fp16 var_347_to_fp16 = const()[name = string("op_347_to_fp16"), val = fp16(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1]> var_348_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_347_to_fp16)[name = string("op_348_cast_fp16")];
+            tensor<int32, [4]> var_349 = const()[name = string("op_349"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 448]> var_350_cast_fp16 = reshape(shape = var_349, x = key_cast_fp16)[name = string("op_350_cast_fp16")];
+            bool mh_w_9_transpose_x_0 = const()[name = string("mh_w_9_transpose_x_0"), val = bool(true)];
+            bool mh_w_9_transpose_y_0 = const()[name = string("mh_w_9_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 20, 1, 448]> mh_w_9_cast_fp16 = matmul(transpose_x = mh_w_9_transpose_x_0, transpose_y = mh_w_9_transpose_y_0, x = var_348_cast_fp16, y = var_350_cast_fp16)[name = string("mh_w_9_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 448]> mh_w_11_cast_fp16 = add(x = mh_w_9_cast_fp16, y = var_146_cast_fp16)[name = string("mh_w_11_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 448]> var_358_cast_fp16 = softmax(axis = var_277, x = mh_w_11_cast_fp16)[name = string("op_358_cast_fp16")];
+            tensor<int32, [4]> var_359 = const()[name = string("op_359"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 448]> var_360_cast_fp16 = reshape(shape = var_359, x = value_cast_fp16)[name = string("op_360_cast_fp16")];
+            bool attn_5_transpose_x_0 = const()[name = string("attn_5_transpose_x_0"), val = bool(false)];
+            bool attn_5_transpose_y_0 = const()[name = string("attn_5_transpose_y_0"), val = bool(true)];
+            tensor<fp16, [1, 20, 64, 1]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_360_cast_fp16, y = var_358_cast_fp16)[name = string("attn_5_cast_fp16")];
+            tensor<int32, [4]> var_363 = const()[name = string("op_363"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1]> input_11_cast_fp16 = reshape(shape = var_363, x = attn_5_cast_fp16)[name = string("input_11_cast_fp16")];
+            string obj_31_pad_type_0 = const()[name = string("obj_31_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> obj_31_strides_0 = const()[name = string("obj_31_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> obj_31_pad_0 = const()[name = string("obj_31_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> obj_31_dilations_0 = const()[name = string("obj_31_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 obj_31_groups_0 = const()[name = string("obj_31_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(189687872)))];
+            tensor<fp16, [1280]> layers_1_self_attn_o_proj_bias_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192964736)))];
+            tensor<fp16, [1, 1280, 1, 1]> obj_31_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_bias_to_fp16, dilations = obj_31_dilations_0, groups = obj_31_groups_0, pad = obj_31_pad_0, pad_type = obj_31_pad_type_0, strides = obj_31_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_11_cast_fp16)[name = string("obj_31_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_31_cast_fp16)[name = string("inputs_9_cast_fp16")];
+            tensor<int32, [1]> out_9_axes_0 = const()[name = string("out_9_axes_0"), val = tensor<int32, [1]>([1])];
+            fp16 var_385_to_fp16 = const()[name = string("op_385_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_385_to_fp16, x = inputs_9_cast_fp16)[name = string("out_9_cast_fp16")];
+            tensor<fp16, [1280]> obj_33_gamma_0_to_fp16 = const()[name = string("obj_33_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192967360)))];
+            tensor<fp16, [1280]> obj_33_beta_0_to_fp16 = const()[name = string("obj_33_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192969984)))];
+            fp16 obj_33_epsilon_0_to_fp16 = const()[name = string("obj_33_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> obj_33_cast_fp16 = batch_norm(beta = obj_33_beta_0_to_fp16, epsilon = obj_33_epsilon_0_to_fp16, gamma = obj_33_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_9_cast_fp16)[name = string("obj_33_cast_fp16")];
+            string query_pad_type_0 = const()[name = string("query_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> query_strides_0 = const()[name = string("query_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> query_pad_0 = const()[name = string("query_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> query_dilations_0 = const()[name = string("query_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 query_groups_0 = const()[name = string("query_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_encoder_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192972608)))];
+            tensor<fp16, [1280]> layers_1_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("layers_1_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196249472)))];
+            tensor<fp16, [1, 1280, 1, 1]> query_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_bias_to_fp16, dilations = query_dilations_0, groups = query_groups_0, pad = query_pad_0, pad_type = query_pad_type_0, strides = query_strides_0, weight = layers_1_encoder_attn_q_proj_weight_to_fp16, x = obj_33_cast_fp16)[name = string("query_cast_fp16")];
+            tensor<int32, [4]> var_405 = const()[name = string("op_405"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1]> mh_q_cast_fp16 = reshape(shape = var_405, x = query_cast_fp16)[name = string("mh_q_cast_fp16")];
+            fp16 var_407_to_fp16 = const()[name = string("op_407_to_fp16"), val = fp16(0x1p-3)];
+            tensor<fp16, [1, 20, 64, 1]> var_408_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_407_to_fp16)[name = string("op_408_cast_fp16")];
+            tensor<int32, [4]> var_409 = const()[name = string("op_409"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1536]> var_410_cast_fp16 = reshape(shape = var_409, x = obj_35_cast_fp16)[name = string("op_410_cast_fp16")];
+            bool mh_w_13_transpose_x_0 = const()[name = string("mh_w_13_transpose_x_0"), val = bool(true)];
+            bool mh_w_13_transpose_y_0 = const()[name = string("mh_w_13_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 20, 1, 1536]> mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_408_cast_fp16, y = var_410_cast_fp16)[name = string("mh_w_13_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 1536]> mh_w_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_206_cast_fp16)[name = string("mh_w_cast_fp16")];
+            tensor<fp16, [1, 20, 1, 1536]> obj_41_cast_fp16 = softmax(axis = var_277, x = mh_w_cast_fp16)[name = string("obj_41_cast_fp16")];
+            tensor<int32, [4]> var_419 = const()[name = string("op_419"), val = tensor<int32, [4]>([1, 20, 64, -1])];
+            tensor<fp16, [1, 20, 64, 1536]> var_420_cast_fp16 = reshape(shape = var_419, x = obj_37_cast_fp16)[name = string("op_420_cast_fp16")];
+            bool attn_transpose_x_0 = const()[name = string("attn_transpose_x_0"), val = bool(false)];
+            bool attn_transpose_y_0 = const()[name = string("attn_transpose_y_0"), val = bool(true)];
+            tensor<fp16, [1, 20, 64, 1]> attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_420_cast_fp16, y = obj_41_cast_fp16)[name = string("attn_cast_fp16")];
+            tensor<int32, [4]> var_423 = const()[name = string("op_423"), val = tensor<int32, [4]>([1, 1280, 1, -1])];
+            tensor<fp16, [1, 1280, 1, 1]> input_13_cast_fp16 = reshape(shape = var_423, x = attn_cast_fp16)[name = string("input_13_cast_fp16")];
+            string obj_39_pad_type_0 = const()[name = string("obj_39_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> obj_39_strides_0 = const()[name = string("obj_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> obj_39_pad_0 = const()[name = string("obj_39_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> obj_39_dilations_0 = const()[name = string("obj_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 obj_39_groups_0 = const()[name = string("obj_39_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 1280, 1, 1]> layers_1_encoder_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_encoder_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1280, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196252096)))];
+            tensor<fp16, [1280]> layers_1_encoder_attn_o_proj_bias_to_fp16 = const()[name = string("layers_1_encoder_attn_o_proj_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199528960)))];
+            tensor<fp16, [1, 1280, 1, 1]> obj_39_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_bias_to_fp16, dilations = obj_39_dilations_0, groups = obj_39_groups_0, pad = obj_39_pad_0, pad_type = obj_39_pad_type_0, strides = obj_39_strides_0, weight = layers_1_encoder_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = string("obj_39_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_39_cast_fp16)[name = string("inputs_11_cast_fp16")];
+            tensor<int32, [1]> out_11_axes_0 = const()[name = string("out_11_axes_0"), val = tensor<int32, [1]>([1])];
+            fp16 var_444_to_fp16 = const()[name = string("op_444_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_444_to_fp16, x = inputs_11_cast_fp16)[name = string("out_11_cast_fp16")];
+            tensor<fp16, [1280]> input_15_gamma_0_to_fp16 = const()[name = string("input_15_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199531584)))];
+            tensor<fp16, [1280]> input_15_beta_0_to_fp16 = const()[name = string("input_15_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199534208)))];
+            fp16 input_15_epsilon_0_to_fp16 = const()[name = string("input_15_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> input_15_cast_fp16 = batch_norm(beta = input_15_beta_0_to_fp16, epsilon = input_15_epsilon_0_to_fp16, gamma = input_15_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_11_cast_fp16)[name = string("input_15_cast_fp16")];
+            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
+            tensor<fp16, [5120, 1280, 1, 1]> layers_1_fc1_weight_to_fp16 = const()[name = string("layers_1_fc1_weight_to_fp16"), val = tensor<fp16, [5120, 1280, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199536832)))];
+            tensor<fp16, [5120]> layers_1_fc1_bias_to_fp16 = const()[name = string("layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [5120]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212644096)))];
+            tensor<fp16, [1, 5120, 1, 1]> input_17_cast_fp16 = conv(bias = layers_1_fc1_bias_to_fp16, dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")];
+            string input_mode_0 = const()[name = string("input_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 5120, 1, 1]> input_cast_fp16 = gelu(mode = input_mode_0, x = input_17_cast_fp16)[name = string("input_cast_fp16")];
+            string hidden_states_5_pad_type_0 = const()[name = string("hidden_states_5_pad_type_0"), val = string("valid")];
+            tensor<int32, [2]> hidden_states_5_strides_0 = const()[name = string("hidden_states_5_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> hidden_states_5_pad_0 = const()[name = string("hidden_states_5_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> hidden_states_5_dilations_0 = const()[name = string("hidden_states_5_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            int32 hidden_states_5_groups_0 = const()[name = string("hidden_states_5_groups_0"), val = int32(1)];
+            tensor<fp16, [1280, 5120, 1, 1]> layers_1_fc2_weight_to_fp16 = const()[name = string("layers_1_fc2_weight_to_fp16"), val = tensor<fp16, [1280, 5120, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(212654400)))];
+            tensor<fp16, [1280]> layers_1_fc2_bias_to_fp16 = const()[name = string("layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225761664)))];
+            tensor<fp16, [1, 1280, 1, 1]> hidden_states_5_cast_fp16 = conv(bias = layers_1_fc2_bias_to_fp16, dilations = hidden_states_5_dilations_0, groups = hidden_states_5_groups_0, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = hidden_states_5_strides_0, weight = layers_1_fc2_weight_to_fp16, x = input_cast_fp16)[name = string("hidden_states_5_cast_fp16")];
+            tensor<fp16, [1, 1280, 1, 1]> inputs_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = string("inputs_cast_fp16")];
+            tensor<int32, [1]> out_axes_0 = const()[name = string("out_axes_0"), val = tensor<int32, [1]>([1])];
+            fp16 var_487_to_fp16 = const()[name = string("op_487_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_487_to_fp16, x = inputs_cast_fp16)[name = string("out_cast_fp16")];
+            tensor<fp16, [1280]> hidden_states_gamma_0_to_fp16 = const()[name = string("hidden_states_gamma_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225764288)))];
+            tensor<fp16, [1280]> hidden_states_beta_0_to_fp16 = const()[name = string("hidden_states_beta_0_to_fp16"), val = tensor<fp16, [1280]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225766912)))];
+            fp16 hidden_states_epsilon_0_to_fp16 = const()[name = string("hidden_states_epsilon_0_to_fp16"), val = fp16(0x1.5p-17)];
+            tensor<fp16, [1, 1280, 1, 1]> hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_5_mean_0_to_fp16, variance = obj_5_variance_0_to_fp16, x = out_cast_fp16)[name = string("hidden_states_cast_fp16")];
+            tensor<int32, [1]> var_498_axes_0 = const()[name = string("op_498_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp16, [1, 1280, 1]> var_498_cast_fp16 = squeeze(axes = var_498_axes_0, x = hidden_states_cast_fp16)[name = string("op_498_cast_fp16")];
+            tensor<int32, [3]> var_501_perm_0 = const()[name = string("op_501_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp16, [51866]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [51866]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(225769536)))];
+            tensor<fp16, [1, 1, 1280]> var_501_cast_fp16 = transpose(perm = var_501_perm_0, x = var_498_cast_fp16)[name = string("transpose_0")];
+            tensor<fp16, [1, 1, 51866]> logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_501_cast_fp16)[name = string("linear_0_cast_fp16")];
+            int32 var_505 = const()[name = string("op_505"), val = int32(1)];
+            bool obj_45_interleave_0 = const()[name = string("obj_45_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2560, 1, 1]> key_cache_updates = concat(axis = var_505, interleave = obj_45_interleave_0, values = (current_key_1_cast_fp16, current_key_cast_fp16))[name = string("obj_45_cast_fp16")];
+            int32 var_508 = const()[name = string("op_508"), val = int32(1)];
+            bool obj_47_interleave_0 = const()[name = string("obj_47_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 2560, 1, 1]> value_cache_updates = concat(axis = var_508, interleave = obj_47_interleave_0, values = (current_value_1_cast_fp16, current_value_cast_fp16))[name = string("obj_47_cast_fp16")];
+            tensor<int32, [4]> var_519_begin_0 = const()[name = string("op_519_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_519_end_0 = const()[name = string("op_519_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_519_end_mask_0 = const()[name = string("op_519_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_519_cast_fp16 = slice_by_index(begin = var_519_begin_0, end = var_519_end_0, end_mask = var_519_end_mask_0, x = obj_41_cast_fp16)[name = string("op_519_cast_fp16")];
+            tensor<int32, [4]> var_522_begin_0 = const()[name = string("op_522_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_522_end_0 = const()[name = string("op_522_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_522_end_mask_0 = const()[name = string("op_522_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_522_squeeze_mask_0 = const()[name = string("op_522_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_522_cast_fp16 = slice_by_index(begin = var_522_begin_0, end = var_522_end_0, end_mask = var_522_end_mask_0, squeeze_mask = var_522_squeeze_mask_0, x = var_519_cast_fp16)[name = string("op_522_cast_fp16")];
+            tensor<int32, [4]> var_537_begin_0 = const()[name = string("op_537_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
+            tensor<int32, [4]> var_537_end_0 = const()[name = string("op_537_end_0"), val = tensor<int32, [4]>([1, 2, 1, 1536])];
+            tensor<bool, [4]> var_537_end_mask_0 = const()[name = string("op_537_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_537_cast_fp16 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = obj_41_cast_fp16)[name = string("op_537_cast_fp16")];
+            tensor<int32, [4]> var_540_begin_0 = const()[name = string("op_540_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_540_end_0 = const()[name = string("op_540_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_540_end_mask_0 = const()[name = string("op_540_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_540_squeeze_mask_0 = const()[name = string("op_540_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_540_cast_fp16 = slice_by_index(begin = var_540_begin_0, end = var_540_end_0, end_mask = var_540_end_mask_0, squeeze_mask = var_540_squeeze_mask_0, x = var_537_cast_fp16)[name = string("op_540_cast_fp16")];
+            tensor<int32, [4]> var_555_begin_0 = const()[name = string("op_555_begin_0"), val = tensor<int32, [4]>([0, 2, 0, 0])];
+            tensor<int32, [4]> var_555_end_0 = const()[name = string("op_555_end_0"), val = tensor<int32, [4]>([1, 3, 1, 1536])];
+            tensor<bool, [4]> var_555_end_mask_0 = const()[name = string("op_555_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_555_cast_fp16 = slice_by_index(begin = var_555_begin_0, end = var_555_end_0, end_mask = var_555_end_mask_0, x = obj_41_cast_fp16)[name = string("op_555_cast_fp16")];
+            tensor<int32, [4]> var_558_begin_0 = const()[name = string("op_558_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_558_end_0 = const()[name = string("op_558_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_558_end_mask_0 = const()[name = string("op_558_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_558_squeeze_mask_0 = const()[name = string("op_558_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_558_cast_fp16 = slice_by_index(begin = var_558_begin_0, end = var_558_end_0, end_mask = var_558_end_mask_0, squeeze_mask = var_558_squeeze_mask_0, x = var_555_cast_fp16)[name = string("op_558_cast_fp16")];
+            tensor<int32, [4]> var_573_begin_0 = const()[name = string("op_573_begin_0"), val = tensor<int32, [4]>([0, 3, 0, 0])];
+            tensor<int32, [4]> var_573_end_0 = const()[name = string("op_573_end_0"), val = tensor<int32, [4]>([1, 4, 1, 1536])];
+            tensor<bool, [4]> var_573_end_mask_0 = const()[name = string("op_573_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_573_cast_fp16 = slice_by_index(begin = var_573_begin_0, end = var_573_end_0, end_mask = var_573_end_mask_0, x = obj_41_cast_fp16)[name = string("op_573_cast_fp16")];
+            tensor<int32, [4]> var_576_begin_0 = const()[name = string("op_576_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_576_end_0 = const()[name = string("op_576_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_576_end_mask_0 = const()[name = string("op_576_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_576_squeeze_mask_0 = const()[name = string("op_576_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_576_cast_fp16 = slice_by_index(begin = var_576_begin_0, end = var_576_end_0, end_mask = var_576_end_mask_0, squeeze_mask = var_576_squeeze_mask_0, x = var_573_cast_fp16)[name = string("op_576_cast_fp16")];
+            tensor<int32, [4]> var_591_begin_0 = const()[name = string("op_591_begin_0"), val = tensor<int32, [4]>([0, 4, 0, 0])];
+            tensor<int32, [4]> var_591_end_0 = const()[name = string("op_591_end_0"), val = tensor<int32, [4]>([1, 5, 1, 1536])];
+            tensor<bool, [4]> var_591_end_mask_0 = const()[name = string("op_591_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_591_cast_fp16 = slice_by_index(begin = var_591_begin_0, end = var_591_end_0, end_mask = var_591_end_mask_0, x = obj_41_cast_fp16)[name = string("op_591_cast_fp16")];
+            tensor<int32, [4]> var_594_begin_0 = const()[name = string("op_594_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_594_end_0 = const()[name = string("op_594_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_594_end_mask_0 = const()[name = string("op_594_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_594_squeeze_mask_0 = const()[name = string("op_594_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_594_cast_fp16 = slice_by_index(begin = var_594_begin_0, end = var_594_end_0, end_mask = var_594_end_mask_0, squeeze_mask = var_594_squeeze_mask_0, x = var_591_cast_fp16)[name = string("op_594_cast_fp16")];
+            tensor<int32, [4]> var_609_begin_0 = const()[name = string("op_609_begin_0"), val = tensor<int32, [4]>([0, 5, 0, 0])];
+            tensor<int32, [4]> var_609_end_0 = const()[name = string("op_609_end_0"), val = tensor<int32, [4]>([1, 6, 1, 1536])];
+            tensor<bool, [4]> var_609_end_mask_0 = const()[name = string("op_609_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_609_cast_fp16 = slice_by_index(begin = var_609_begin_0, end = var_609_end_0, end_mask = var_609_end_mask_0, x = obj_41_cast_fp16)[name = string("op_609_cast_fp16")];
+            tensor<int32, [4]> var_612_begin_0 = const()[name = string("op_612_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_612_end_0 = const()[name = string("op_612_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_612_end_mask_0 = const()[name = string("op_612_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_612_squeeze_mask_0 = const()[name = string("op_612_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_612_cast_fp16 = slice_by_index(begin = var_612_begin_0, end = var_612_end_0, end_mask = var_612_end_mask_0, squeeze_mask = var_612_squeeze_mask_0, x = var_609_cast_fp16)[name = string("op_612_cast_fp16")];
+            tensor<int32, [4]> var_627_begin_0 = const()[name = string("op_627_begin_0"), val = tensor<int32, [4]>([0, 6, 0, 0])];
+            tensor<int32, [4]> var_627_end_0 = const()[name = string("op_627_end_0"), val = tensor<int32, [4]>([1, 7, 1, 1536])];
+            tensor<bool, [4]> var_627_end_mask_0 = const()[name = string("op_627_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_627_cast_fp16 = slice_by_index(begin = var_627_begin_0, end = var_627_end_0, end_mask = var_627_end_mask_0, x = obj_41_cast_fp16)[name = string("op_627_cast_fp16")];
+            tensor<int32, [4]> var_630_begin_0 = const()[name = string("op_630_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_630_end_0 = const()[name = string("op_630_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_630_end_mask_0 = const()[name = string("op_630_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_630_squeeze_mask_0 = const()[name = string("op_630_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_630_cast_fp16 = slice_by_index(begin = var_630_begin_0, end = var_630_end_0, end_mask = var_630_end_mask_0, squeeze_mask = var_630_squeeze_mask_0, x = var_627_cast_fp16)[name = string("op_630_cast_fp16")];
+            tensor<int32, [4]> var_645_begin_0 = const()[name = string("op_645_begin_0"), val = tensor<int32, [4]>([0, 7, 0, 0])];
+            tensor<int32, [4]> var_645_end_0 = const()[name = string("op_645_end_0"), val = tensor<int32, [4]>([1, 8, 1, 1536])];
+            tensor<bool, [4]> var_645_end_mask_0 = const()[name = string("op_645_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_645_cast_fp16 = slice_by_index(begin = var_645_begin_0, end = var_645_end_0, end_mask = var_645_end_mask_0, x = obj_41_cast_fp16)[name = string("op_645_cast_fp16")];
+            tensor<int32, [4]> var_648_begin_0 = const()[name = string("op_648_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_648_end_0 = const()[name = string("op_648_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_648_end_mask_0 = const()[name = string("op_648_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_648_squeeze_mask_0 = const()[name = string("op_648_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_648_cast_fp16 = slice_by_index(begin = var_648_begin_0, end = var_648_end_0, end_mask = var_648_end_mask_0, squeeze_mask = var_648_squeeze_mask_0, x = var_645_cast_fp16)[name = string("op_648_cast_fp16")];
+            tensor<int32, [4]> var_663_begin_0 = const()[name = string("op_663_begin_0"), val = tensor<int32, [4]>([0, 8, 0, 0])];
+            tensor<int32, [4]> var_663_end_0 = const()[name = string("op_663_end_0"), val = tensor<int32, [4]>([1, 9, 1, 1536])];
+            tensor<bool, [4]> var_663_end_mask_0 = const()[name = string("op_663_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_663_cast_fp16 = slice_by_index(begin = var_663_begin_0, end = var_663_end_0, end_mask = var_663_end_mask_0, x = obj_41_cast_fp16)[name = string("op_663_cast_fp16")];
+            tensor<int32, [4]> var_666_begin_0 = const()[name = string("op_666_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_666_end_0 = const()[name = string("op_666_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_666_end_mask_0 = const()[name = string("op_666_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_666_squeeze_mask_0 = const()[name = string("op_666_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_666_cast_fp16 = slice_by_index(begin = var_666_begin_0, end = var_666_end_0, end_mask = var_666_end_mask_0, squeeze_mask = var_666_squeeze_mask_0, x = var_663_cast_fp16)[name = string("op_666_cast_fp16")];
+            tensor<int32, [4]> var_681_begin_0 = const()[name = string("op_681_begin_0"), val = tensor<int32, [4]>([0, 9, 0, 0])];
+            tensor<int32, [4]> var_681_end_0 = const()[name = string("op_681_end_0"), val = tensor<int32, [4]>([1, 10, 1, 1536])];
+            tensor<bool, [4]> var_681_end_mask_0 = const()[name = string("op_681_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_681_cast_fp16 = slice_by_index(begin = var_681_begin_0, end = var_681_end_0, end_mask = var_681_end_mask_0, x = obj_41_cast_fp16)[name = string("op_681_cast_fp16")];
+            tensor<int32, [4]> var_684_begin_0 = const()[name = string("op_684_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_684_end_0 = const()[name = string("op_684_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_684_end_mask_0 = const()[name = string("op_684_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_684_squeeze_mask_0 = const()[name = string("op_684_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_684_cast_fp16 = slice_by_index(begin = var_684_begin_0, end = var_684_end_0, end_mask = var_684_end_mask_0, squeeze_mask = var_684_squeeze_mask_0, x = var_681_cast_fp16)[name = string("op_684_cast_fp16")];
+            tensor<int32, [4]> var_699_begin_0 = const()[name = string("op_699_begin_0"), val = tensor<int32, [4]>([0, 10, 0, 0])];
+            tensor<int32, [4]> var_699_end_0 = const()[name = string("op_699_end_0"), val = tensor<int32, [4]>([1, 11, 1, 1536])];
+            tensor<bool, [4]> var_699_end_mask_0 = const()[name = string("op_699_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_699_cast_fp16 = slice_by_index(begin = var_699_begin_0, end = var_699_end_0, end_mask = var_699_end_mask_0, x = obj_41_cast_fp16)[name = string("op_699_cast_fp16")];
+            tensor<int32, [4]> var_702_begin_0 = const()[name = string("op_702_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_702_end_0 = const()[name = string("op_702_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_702_end_mask_0 = const()[name = string("op_702_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_702_squeeze_mask_0 = const()[name = string("op_702_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_702_cast_fp16 = slice_by_index(begin = var_702_begin_0, end = var_702_end_0, end_mask = var_702_end_mask_0, squeeze_mask = var_702_squeeze_mask_0, x = var_699_cast_fp16)[name = string("op_702_cast_fp16")];
+            tensor<int32, [4]> var_717_begin_0 = const()[name = string("op_717_begin_0"), val = tensor<int32, [4]>([0, 11, 0, 0])];
+            tensor<int32, [4]> var_717_end_0 = const()[name = string("op_717_end_0"), val = tensor<int32, [4]>([1, 12, 1, 1536])];
+            tensor<bool, [4]> var_717_end_mask_0 = const()[name = string("op_717_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_717_cast_fp16 = slice_by_index(begin = var_717_begin_0, end = var_717_end_0, end_mask = var_717_end_mask_0, x = obj_41_cast_fp16)[name = string("op_717_cast_fp16")];
+            tensor<int32, [4]> var_720_begin_0 = const()[name = string("op_720_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_720_end_0 = const()[name = string("op_720_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_720_end_mask_0 = const()[name = string("op_720_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_720_squeeze_mask_0 = const()[name = string("op_720_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_720_cast_fp16 = slice_by_index(begin = var_720_begin_0, end = var_720_end_0, end_mask = var_720_end_mask_0, squeeze_mask = var_720_squeeze_mask_0, x = var_717_cast_fp16)[name = string("op_720_cast_fp16")];
+            tensor<int32, [4]> var_735_begin_0 = const()[name = string("op_735_begin_0"), val = tensor<int32, [4]>([0, 12, 0, 0])];
+            tensor<int32, [4]> var_735_end_0 = const()[name = string("op_735_end_0"), val = tensor<int32, [4]>([1, 13, 1, 1536])];
+            tensor<bool, [4]> var_735_end_mask_0 = const()[name = string("op_735_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_735_cast_fp16 = slice_by_index(begin = var_735_begin_0, end = var_735_end_0, end_mask = var_735_end_mask_0, x = obj_41_cast_fp16)[name = string("op_735_cast_fp16")];
+            tensor<int32, [4]> var_738_begin_0 = const()[name = string("op_738_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_738_end_0 = const()[name = string("op_738_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_738_end_mask_0 = const()[name = string("op_738_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_738_squeeze_mask_0 = const()[name = string("op_738_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_738_cast_fp16 = slice_by_index(begin = var_738_begin_0, end = var_738_end_0, end_mask = var_738_end_mask_0, squeeze_mask = var_738_squeeze_mask_0, x = var_735_cast_fp16)[name = string("op_738_cast_fp16")];
+            tensor<int32, [4]> var_753_begin_0 = const()[name = string("op_753_begin_0"), val = tensor<int32, [4]>([0, 13, 0, 0])];
+            tensor<int32, [4]> var_753_end_0 = const()[name = string("op_753_end_0"), val = tensor<int32, [4]>([1, 14, 1, 1536])];
+            tensor<bool, [4]> var_753_end_mask_0 = const()[name = string("op_753_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_753_cast_fp16 = slice_by_index(begin = var_753_begin_0, end = var_753_end_0, end_mask = var_753_end_mask_0, x = obj_41_cast_fp16)[name = string("op_753_cast_fp16")];
+            tensor<int32, [4]> var_756_begin_0 = const()[name = string("op_756_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_756_end_0 = const()[name = string("op_756_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_756_end_mask_0 = const()[name = string("op_756_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_756_squeeze_mask_0 = const()[name = string("op_756_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_756_cast_fp16 = slice_by_index(begin = var_756_begin_0, end = var_756_end_0, end_mask = var_756_end_mask_0, squeeze_mask = var_756_squeeze_mask_0, x = var_753_cast_fp16)[name = string("op_756_cast_fp16")];
+            tensor<int32, [4]> var_771_begin_0 = const()[name = string("op_771_begin_0"), val = tensor<int32, [4]>([0, 14, 0, 0])];
+            tensor<int32, [4]> var_771_end_0 = const()[name = string("op_771_end_0"), val = tensor<int32, [4]>([1, 15, 1, 1536])];
+            tensor<bool, [4]> var_771_end_mask_0 = const()[name = string("op_771_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_771_cast_fp16 = slice_by_index(begin = var_771_begin_0, end = var_771_end_0, end_mask = var_771_end_mask_0, x = obj_41_cast_fp16)[name = string("op_771_cast_fp16")];
+            tensor<int32, [4]> var_774_begin_0 = const()[name = string("op_774_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_774_end_0 = const()[name = string("op_774_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_774_end_mask_0 = const()[name = string("op_774_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_774_squeeze_mask_0 = const()[name = string("op_774_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_774_cast_fp16 = slice_by_index(begin = var_774_begin_0, end = var_774_end_0, end_mask = var_774_end_mask_0, squeeze_mask = var_774_squeeze_mask_0, x = var_771_cast_fp16)[name = string("op_774_cast_fp16")];
+            tensor<int32, [4]> var_789_begin_0 = const()[name = string("op_789_begin_0"), val = tensor<int32, [4]>([0, 15, 0, 0])];
+            tensor<int32, [4]> var_789_end_0 = const()[name = string("op_789_end_0"), val = tensor<int32, [4]>([1, 16, 1, 1536])];
+            tensor<bool, [4]> var_789_end_mask_0 = const()[name = string("op_789_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_789_cast_fp16 = slice_by_index(begin = var_789_begin_0, end = var_789_end_0, end_mask = var_789_end_mask_0, x = obj_41_cast_fp16)[name = string("op_789_cast_fp16")];
+            tensor<int32, [4]> var_792_begin_0 = const()[name = string("op_792_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_792_end_0 = const()[name = string("op_792_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_792_end_mask_0 = const()[name = string("op_792_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_792_squeeze_mask_0 = const()[name = string("op_792_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_792_cast_fp16 = slice_by_index(begin = var_792_begin_0, end = var_792_end_0, end_mask = var_792_end_mask_0, squeeze_mask = var_792_squeeze_mask_0, x = var_789_cast_fp16)[name = string("op_792_cast_fp16")];
+            tensor<int32, [4]> var_807_begin_0 = const()[name = string("op_807_begin_0"), val = tensor<int32, [4]>([0, 16, 0, 0])];
+            tensor<int32, [4]> var_807_end_0 = const()[name = string("op_807_end_0"), val = tensor<int32, [4]>([1, 17, 1, 1536])];
+            tensor<bool, [4]> var_807_end_mask_0 = const()[name = string("op_807_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_807_cast_fp16 = slice_by_index(begin = var_807_begin_0, end = var_807_end_0, end_mask = var_807_end_mask_0, x = obj_41_cast_fp16)[name = string("op_807_cast_fp16")];
+            tensor<int32, [4]> var_810_begin_0 = const()[name = string("op_810_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_810_end_0 = const()[name = string("op_810_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_810_end_mask_0 = const()[name = string("op_810_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_810_squeeze_mask_0 = const()[name = string("op_810_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_810_cast_fp16 = slice_by_index(begin = var_810_begin_0, end = var_810_end_0, end_mask = var_810_end_mask_0, squeeze_mask = var_810_squeeze_mask_0, x = var_807_cast_fp16)[name = string("op_810_cast_fp16")];
+            tensor<int32, [4]> var_825_begin_0 = const()[name = string("op_825_begin_0"), val = tensor<int32, [4]>([0, 17, 0, 0])];
+            tensor<int32, [4]> var_825_end_0 = const()[name = string("op_825_end_0"), val = tensor<int32, [4]>([1, 18, 1, 1536])];
+            tensor<bool, [4]> var_825_end_mask_0 = const()[name = string("op_825_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_825_cast_fp16 = slice_by_index(begin = var_825_begin_0, end = var_825_end_0, end_mask = var_825_end_mask_0, x = obj_41_cast_fp16)[name = string("op_825_cast_fp16")];
+            tensor<int32, [4]> var_828_begin_0 = const()[name = string("op_828_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_828_end_0 = const()[name = string("op_828_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_828_end_mask_0 = const()[name = string("op_828_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_828_squeeze_mask_0 = const()[name = string("op_828_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_828_cast_fp16 = slice_by_index(begin = var_828_begin_0, end = var_828_end_0, end_mask = var_828_end_mask_0, squeeze_mask = var_828_squeeze_mask_0, x = var_825_cast_fp16)[name = string("op_828_cast_fp16")];
+            tensor<int32, [4]> var_843_begin_0 = const()[name = string("op_843_begin_0"), val = tensor<int32, [4]>([0, 18, 0, 0])];
+            tensor<int32, [4]> var_843_end_0 = const()[name = string("op_843_end_0"), val = tensor<int32, [4]>([1, 19, 1, 1536])];
+            tensor<bool, [4]> var_843_end_mask_0 = const()[name = string("op_843_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_843_cast_fp16 = slice_by_index(begin = var_843_begin_0, end = var_843_end_0, end_mask = var_843_end_mask_0, x = obj_41_cast_fp16)[name = string("op_843_cast_fp16")];
+            tensor<int32, [4]> var_846_begin_0 = const()[name = string("op_846_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_846_end_0 = const()[name = string("op_846_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_846_end_mask_0 = const()[name = string("op_846_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_846_squeeze_mask_0 = const()[name = string("op_846_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_846_cast_fp16 = slice_by_index(begin = var_846_begin_0, end = var_846_end_0, end_mask = var_846_end_mask_0, squeeze_mask = var_846_squeeze_mask_0, x = var_843_cast_fp16)[name = string("op_846_cast_fp16")];
+            tensor<int32, [4]> var_861_begin_0 = const()[name = string("op_861_begin_0"), val = tensor<int32, [4]>([0, 19, 0, 0])];
+            tensor<int32, [4]> var_861_end_0 = const()[name = string("op_861_end_0"), val = tensor<int32, [4]>([1, 20, 1, 1536])];
+            tensor<bool, [4]> var_861_end_mask_0 = const()[name = string("op_861_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
+            tensor<fp16, [1, 1, 1, 1536]> var_861_cast_fp16 = slice_by_index(begin = var_861_begin_0, end = var_861_end_0, end_mask = var_861_end_mask_0, x = obj_41_cast_fp16)[name = string("op_861_cast_fp16")];
+            tensor<int32, [4]> var_864_begin_0 = const()[name = string("op_864_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> var_864_end_0 = const()[name = string("op_864_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1536])];
+            tensor<bool, [4]> var_864_end_mask_0 = const()[name = string("op_864_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
+            tensor<bool, [4]> var_864_squeeze_mask_0 = const()[name = string("op_864_squeeze_mask_0"), val = tensor<bool, [4]>([false, false, true, false])];
+            tensor<fp16, [1, 1, 1536]> var_864_cast_fp16 = slice_by_index(begin = var_864_begin_0, end = var_864_end_0, end_mask = var_864_end_mask_0, squeeze_mask = var_864_squeeze_mask_0, x = var_861_cast_fp16)[name = string("op_864_cast_fp16")];
+            int32 var_871 = const()[name = string("op_871"), val = int32(1)];
+            bool var_872_interleave_0 = const()[name = string("op_872_interleave_0"), val = bool(false)];
+            tensor<fp16, [1, 20, 1536]> var_872_cast_fp16 = concat(axis = var_871, interleave = var_872_interleave_0, values = (var_522_cast_fp16, var_540_cast_fp16, var_558_cast_fp16, var_576_cast_fp16, var_594_cast_fp16, var_612_cast_fp16, var_630_cast_fp16, var_648_cast_fp16, var_666_cast_fp16, var_684_cast_fp16, var_702_cast_fp16, var_720_cast_fp16, var_738_cast_fp16, var_756_cast_fp16, var_774_cast_fp16, var_792_cast_fp16, var_810_cast_fp16, var_828_cast_fp16, var_846_cast_fp16, var_864_cast_fp16))[name = string("op_872_cast_fp16")];
+            bool var_875 = const()[name = string("op_875"), val = bool(false)];
+            tensor<int32, [1]> obj_axes_0 = const()[name = string("obj_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1536]> alignment_heads_weights = reduce_mean(axes = obj_axes_0, keep_dims = var_875, x = var_872_cast_fp16)[name = string("obj_cast_fp16")];
+        } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights);
+}

distil-whisper_distil-large-v3_turbo/TextDecoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c1afaacaec2fac64e8867d758742347e10c849fdbf81c8761344b5c56a55b5d
+size 225873332

distil-whisper_distil-large-v3_turbo/config.json

@@ -0,0 +1 @@
+{"_name_or_path": "./distil-large-v3", "activation_dropout": 0.0, "activation_function": "gelu", "apply_spec_augment": false, "architectures": ["WhisperForConditionalGeneration"], "attention_dropout": 0.0, "begin_suppress_tokens": [220, 50257], "bos_token_id": 50257, "classifier_proj_size": 256, "d_model": 1280, "decoder_attention_heads": 20, "decoder_ffn_dim": 5120, "decoder_layerdrop": 0.0, "decoder_layers": 2, "decoder_start_token_id": 50258, "dropout": 0.0, "encoder_attention_heads": 20, "encoder_ffn_dim": 5120, "encoder_layerdrop": 0.0, "encoder_layers": 32, "eos_token_id": 50257, "init_std": 0.02, "is_encoder_decoder": true, "mask_feature_length": 10, "mask_feature_min_masks": 0, "mask_feature_prob": 0.0, "mask_time_length": 10, "mask_time_min_masks": 2, "mask_time_prob": 0.05, "max_length": 448, "max_source_positions": 1500, "max_target_positions": 448, "median_filter_width": 7, "model_type": "whisper", "num_hidden_layers": 32, "num_mel_bins": 128, "pad_token_id": 50256, "scale_embedding": false, "torch_dtype": "float16", "transformers_version": "4.38.0.dev0", "use_cache": true, "use_weighted_layer_sum": false, "vocab_size": 51866}

distil-whisper_distil-large-v3_turbo/generation_config.json

@@ -0,0 +1 @@
+{"alignment_heads": [[7, 0], [10, 17], [12, 18], [13, 12], [16, 1], [17, 14], [19, 11], [21, 4], [24, 1], [25, 6]], "begin_suppress_tokens": [220, 50257], "bos_token_id": 50257, "decoder_start_token_id": 50258, "eos_token_id": 50257, "forced_decoder_ids": [[1, null], [2, 50360]], "is_multilingual": true, "lang_to_id": {"<|af|>": 50327, "<|am|>": 50334, "<|ar|>": 50272, "<|as|>": 50350, "<|az|>": 50304, "<|ba|>": 50355, "<|be|>": 50330, "<|bg|>": 50292, "<|bn|>": 50302, "<|bo|>": 50347, "<|br|>": 50309, "<|bs|>": 50315, "<|ca|>": 50270, "<|cs|>": 50283, "<|cy|>": 50297, "<|da|>": 50285, "<|de|>": 50261, "<|el|>": 50281, "<|en|>": 50259, "<|es|>": 50262, "<|et|>": 50307, "<|eu|>": 50310, "<|fa|>": 50300, "<|fi|>": 50277, "<|fo|>": 50338, "<|fr|>": 50265, "<|gl|>": 50319, "<|gu|>": 50333, "<|haw|>": 50352, "<|ha|>": 50354, "<|he|>": 50279, "<|hi|>": 50276, "<|hr|>": 50291, "<|ht|>": 50339, "<|hu|>": 50286, "<|hy|>": 50312, "<|id|>": 50275, "<|is|>": 50311, "<|it|>": 50274, "<|ja|>": 50266, "<|jw|>": 50356, "<|ka|>": 50329, "<|kk|>": 50316, "<|km|>": 50323, "<|kn|>": 50306, "<|ko|>": 50264, "<|la|>": 50294, "<|lb|>": 50345, "<|ln|>": 50353, "<|lo|>": 50336, "<|lt|>": 50293, "<|lv|>": 50301, "<|mg|>": 50349, "<|mi|>": 50295, "<|mk|>": 50308, "<|ml|>": 50296, "<|mn|>": 50314, "<|mr|>": 50320, "<|ms|>": 50282, "<|mt|>": 50343, "<|my|>": 50346, "<|ne|>": 50313, "<|nl|>": 50271, "<|nn|>": 50342, "<|no|>": 50288, "<|oc|>": 50328, "<|pa|>": 50321, "<|pl|>": 50269, "<|ps|>": 50340, "<|pt|>": 50267, "<|ro|>": 50284, "<|ru|>": 50263, "<|sa|>": 50344, "<|sd|>": 50332, "<|si|>": 50322, "<|sk|>": 50298, "<|sl|>": 50305, "<|sn|>": 50324, "<|so|>": 50326, "<|sq|>": 50317, "<|sr|>": 50303, "<|su|>": 50357, "<|sv|>": 50273, "<|sw|>": 50318, "<|ta|>": 50287, "<|te|>": 50299, "<|tg|>": 50331, "<|th|>": 50289, "<|tk|>": 50341, "<|tl|>": 50348, "<|tr|>": 50268, "<|tt|>": 50351, "<|uk|>": 50280, "<|ur|>": 50290, "<|uz|>": 50337, "<|vi|>": 50278, "<|yi|>": 50335, "<|yo|>": 50325, "<|yue|>": 50358, "<|zh|>": 50260}, "language": "<|en|>", "max_initial_timestamp_index": 50, "max_length": 448, "no_timestamps_token_id": 50364, "pad_token_id": 50257, "prev_sot_token_id": 50362, "return_timestamps": false, "suppress_tokens": [1, 2, 7, 8, 9, 10, 14, 25, 26, 27, 28, 29, 31, 58, 59, 60, 61, 62, 63, 90, 91, 92, 93, 359, 503, 522, 542, 873, 893, 902, 918, 922, 931, 1350, 1853, 1982, 2460, 2627, 3246, 3253, 3268, 3536, 3846, 3961, 4183, 4667, 6585, 6647, 7273, 9061, 9383, 10428, 10929, 11938, 12033, 12331, 12562, 13793, 14157, 14635, 15265, 15618, 16553, 16604, 18362, 18956, 20075, 21675, 22520, 26130, 26161, 26435, 28279, 29464, 31650, 32302, 32470, 36865, 42863, 47425, 49870, 50254, 50258, 50359, 50360, 50361, 50362, 50363], "task": "transcribe", "task_to_id": {"transcribe": 50360, "translate": 50359}, "transformers_version": "4.38.0.dev0"}

openai_whisper-base.en/AudioEncoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:384d19c754b6ca6a7ad6dd457406dd9c9de44e43034cbfaf3f343e0278e43ac9
+size 243

openai_whisper-base.en/AudioEncoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a536e74da525305d998542cdad99de17f18771834664969738d6fa2ab99fd115
+size 433

openai_whisper-base.en/AudioEncoder.mlmodelc/metadata.json

@@ -0,0 +1,91 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 512 × 1 × 1500)",
+        "shortDescription" : "",
+        "shape" : "[1, 512, 1, 1500]",
+        "name" : "encoder_output_embeds",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 6 × 512 × 1 × 1536)",
+        "shortDescription" : "",
+        "shape" : "[6, 512, 1, 1536]",
+        "name" : "encoder_attn_key_cache",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 6 × 512 × 1 × 1536)",
+        "shortDescription" : "",
+        "shape" : "[6, 512, 1, 1536]",
+        "name" : "encoder_attn_value_cache",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Pad" : 2,
+      "Ios18.batchNorm" : 13,
+      "Ios18.conv" : 50,
+      "Ios18.gelu" : 8,
+      "Ios18.concat" : 56,
+      "Ios16.einsum" : 384,
+      "Ios18.add" : 13,
+      "Ios18.softmax" : 192,
+      "Ios18.sliceByIndex" : 336,
+      "Ios18.layerNorm" : 13,
+      "Ios18.transpose" : 6,
+      "Ios18.mul" : 192
+    },
+    "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.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.5.1",
+      "com.github.apple.coremltools.version" : "8.0"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 80 × 1 × 3000)",
+        "shortDescription" : "",
+        "shape" : "[1, 80, 1, 3000]",
+        "name" : "melspectrogram_features",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "AudioEncoderStateful",
+    "method" : "predict"
+  }
+]

openai_whisper-base.en/AudioEncoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf7338475244e4dfd0df94300cf8449d6df39c592d5f9488360b59265a9cc80d
+size 47488384

openai_whisper-base.en/LICENSE_NOTICE.txt

@@ -0,0 +1,7 @@
+Argmax proprietary and confidential. Under NDA.
+
+Copyright 2024 Argmax, Inc. All rights reserved.
+
+Unauthorized access, copying, use, distribution, and or commercialization of this file, via any medium or means is strictly prohibited.
+
+Please contact Argmax for licensing information at [email protected].

openai_whisper-base.en/MelSpectrogram.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:efc05e563ee0c556e3f578e04be5fb67b4e7520124403f2561f39102f0f2b33d
+size 243

openai_whisper-base.en/MelSpectrogram.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4ef11ea703011eab03287ec661f999e19c2c78cf67d531b5e6afa02e18f913d
+size 328

openai_whisper-base.en/MelSpectrogram.mlmodelc/metadata.json

@@ -0,0 +1,74 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 80 × 1 × 3000)",
+        "shortDescription" : "",
+        "shape" : "[1, 80, 1, 3000]",
+        "name" : "melspectrogram_features",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.mul" : 2,
+      "Ios18.square" : 2,
+      "Ios18.conv" : 2,
+      "Ios18.matmul" : 1,
+      "Ios18.expandDims" : 4,
+      "Ios18.sub" : 1,
+      "Ios18.log" : 1,
+      "Ios18.add" : 3,
+      "Ios18.sliceByIndex" : 1,
+      "Ios18.maximum" : 1,
+      "Ios18.squeeze" : 2,
+      "Ios18.reshape" : 2,
+      "Ios16.reduceMax" : 1,
+      "Identity" : 1,
+      "Pad" : 1
+    },
+    "computePrecision" : "Mixed (Float16, Float32, 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.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.source" : "torch==2.5.1",
+      "com.github.apple.coremltools.version" : "8.0"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 480000)",
+        "shortDescription" : "",
+        "shape" : "[480000]",
+        "name" : "audio",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "MelSpectrogram",
+    "method" : "predict"
+  }
+]