Upload folder using huggingface_hub

jina_coreml_manifest.json

@@ -0,0 +1,28 @@
+{
+  "format_version": 1,
+  "source_model": "jinaai/jina-embeddings-v2-small-en",
+  "conversion": "Core ML MLProgram FP16 static-shape sequence buckets",
+  "minimum_macos": "14.0",
+  "preferred_compute_units": "cpuAndNeuralEngine",
+  "input": "input_ids",
+  "input_dtype": "int32",
+  "output": "embedding",
+  "output_dimensions": 512,
+  "artifacts": [
+    {
+      "sequence_length": 128,
+      "artifact": "jina_embedding_128.mlmodelc",
+      "sha256_tree": "5e8fd8d4edd9d2fdb832d0688b59afc15a47c592bc2b3050298437a88dc72721"
+    },
+    {
+      "sequence_length": 256,
+      "artifact": "jina_embedding_256.mlmodelc",
+      "sha256_tree": "dd003d5c6ac01f7ae06e7c46293820cac449609f997caeacf5206052d3ef8072"
+    },
+    {
+      "sequence_length": 512,
+      "artifact": "jina_embedding.mlmodelc",
+      "sha256_tree": "c4b2f915eb53dc2c596603ec02f71275e31ad507208082acb52431fd3127937f"
+    }
+  ]
+}

jina_embedding.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:23f21dc172e5c0428832d71f335976ab260c5a2d9e3fb28cf61959ace7786004
+size 243

jina_embedding.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a379f66d023f68191f3beafde7450fd1ef6f84bdc76af3a663c22663aedb581
+size 318

jina_embedding.mlmodelc/metadata.json

@@ -0,0 +1,80 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 512)",
+        "shortDescription" : "",
+        "shape" : "[1, 512]",
+        "name" : "embedding",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.mul" : 10,
+      "Ios17.notEqual" : 1,
+      "Ios17.reshape" : 16,
+      "Ios16.softmax" : 4,
+      "Ios17.matmul" : 8,
+      "Ios17.maximum" : 1,
+      "Ios17.transpose" : 16,
+      "Ios17.expandDims" : 3,
+      "Ios17.add" : 17,
+      "Tile" : 1,
+      "Ios17.sliceByIndex" : 8,
+      "Ios16.reduceSum" : 2,
+      "Ios17.gather" : 1,
+      "Ios17.layerNorm" : 9,
+      "Ios17.clip" : 1,
+      "Ios17.sub" : 1,
+      "Ios16.gelu" : 4,
+      "Ios17.cast" : 4,
+      "Ios17.realDiv" : 2,
+      "Ios17.linear" : 24,
+      "Ios16.reduceL2Norm" : 1
+    },
+    "computePrecision" : "Mixed (Float16, Int16, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.source_dialect" : "TorchScript",
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.5.0"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 512)",
+        "shortDescription" : "",
+        "shape" : "[1, 512]",
+        "name" : "input_ids",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "jina_embedding_ane",
+    "method" : "predict"
+  }
+]

jina_embedding.mlmodelc/model.mil

@@ -0,0 +1,320 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.5.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<int32, [1, 512]> input_ids) {
+            tensor<int32, []> var_3 = const()[name = tensor<string, []>("op_3"), val = tensor<int32, []>(0)];
+            tensor<bool, [1, 512]> var_4 = not_equal(x = input_ids, y = var_3)[name = tensor<string, []>("op_4")];
+            tensor<string, []> cast_0_dtype_0 = const()[name = tensor<string, []>("cast_0_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, []> var_16 = const()[name = tensor<string, []>("op_16"), val = tensor<int32, []>(-1)];
+            tensor<int32, [1]> var_41_axes_0 = const()[name = tensor<string, []>("op_41_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 512]> cast_0 = cast(dtype = cast_0_dtype_0, x = var_4)[name = tensor<string, []>("cast_37")];
+            tensor<int32, [1, 1, 512]> var_41 = expand_dims(axes = var_41_axes_0, x = cast_0)[name = tensor<string, []>("op_41")];
+            tensor<int32, [1]> var_42_axes_0 = const()[name = tensor<string, []>("op_42_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<int32, [1, 1, 1, 512]> var_42 = expand_dims(axes = var_42_axes_0, x = var_41)[name = tensor<string, []>("op_42")];
+            tensor<fp16, []> var_21_to_fp16 = const()[name = tensor<string, []>("op_21_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<string, []> cast_1_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 1, 1, 512]> var_42_to_fp16 = cast(dtype = cast_1_to_fp16_dtype_0, x = var_42)[name = tensor<string, []>("cast_36")];
+            tensor<fp16, [1, 1, 1, 512]> var_45_cast_fp16 = sub(x = var_21_to_fp16, y = var_42_to_fp16)[name = tensor<string, []>("op_45_cast_fp16")];
+            tensor<fp16, []> var_46_to_fp16 = const()[name = tensor<string, []>("op_46_to_fp16"), val = tensor<fp16, []>(-0x1.d4cp+14)];
+            tensor<fp16, [1, 1, 1, 512]> attention_mask_cast_fp16 = mul(x = var_45_cast_fp16, y = var_46_to_fp16)[name = tensor<string, []>("attention_mask_cast_fp16")];
+            tensor<int32, []> inputs_embeds_axis_0 = const()[name = tensor<string, []>("inputs_embeds_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> inputs_embeds_batch_dims_0 = const()[name = tensor<string, []>("inputs_embeds_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> inputs_embeds_validate_indices_0 = const()[name = tensor<string, []>("inputs_embeds_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [30528, 512]> m_embeddings_word_embeddings_weight_to_fp16 = const()[name = tensor<string, []>("m_embeddings_word_embeddings_weight_to_fp16"), val = tensor<fp16, [30528, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<string, []> input_ids_to_int16_dtype_0 = const()[name = tensor<string, []>("input_ids_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 512]> input_ids_to_int16 = cast(dtype = input_ids_to_int16_dtype_0, x = input_ids)[name = tensor<string, []>("cast_35")];
+            tensor<fp16, [1, 512, 512]> inputs_embeds_cast_fp16_cast_uint16 = gather(axis = inputs_embeds_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = input_ids_to_int16, validate_indices = inputs_embeds_validate_indices_0, x = m_embeddings_word_embeddings_weight_to_fp16)[name = tensor<string, []>("inputs_embeds_cast_fp16_cast_uint16")];
+            tensor<fp16, [1, 512, 512]> token_type_embeddings_1_to_fp16 = const()[name = tensor<string, []>("token_type_embeddings_1_to_fp16"), val = tensor<fp16, [1, 512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31260800)))];
+            tensor<fp16, [1, 512, 512]> input_3_cast_fp16 = add(x = inputs_embeds_cast_fp16_cast_uint16, y = token_type_embeddings_1_to_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<int32, [1]> input_5_axes_0 = const()[name = tensor<string, []>("input_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_embeddings_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_embeddings_LayerNorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31785152)))];
+            tensor<fp16, [512]> m_embeddings_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_embeddings_LayerNorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31786240)))];
+            tensor<fp16, []> var_18_to_fp16 = const()[name = tensor<string, []>("op_18_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [1, 512, 512]> input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = m_embeddings_LayerNorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_embeddings_LayerNorm_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_0_attention_self_query_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_query_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31787328)))];
+            tensor<fp16, [512]> m_encoder_layer_0_attention_self_query_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_query_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32311680)))];
+            tensor<fp16, [1, 512, 512]> linear_0_cast_fp16 = linear(bias = m_encoder_layer_0_attention_self_query_bias_to_fp16, weight = m_encoder_layer_0_attention_self_query_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_0_attention_self_key_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_key_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32312768)))];
+            tensor<fp16, [512]> m_encoder_layer_0_attention_self_key_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_key_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32837120)))];
+            tensor<fp16, [1, 512, 512]> linear_1_cast_fp16 = linear(bias = m_encoder_layer_0_attention_self_key_bias_to_fp16, weight = m_encoder_layer_0_attention_self_key_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<int32, [4]> var_90 = const()[name = tensor<string, []>("op_90"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_3_cast_fp16 = reshape(shape = var_90, x = linear_1_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_0_attention_self_value_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_value_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32838208)))];
+            tensor<fp16, [512]> m_encoder_layer_0_attention_self_value_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_value_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33362560)))];
+            tensor<fp16, [1, 512, 512]> linear_2_cast_fp16 = linear(bias = m_encoder_layer_0_attention_self_value_bias_to_fp16, weight = m_encoder_layer_0_attention_self_value_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<int32, [4]> var_99 = const()[name = tensor<string, []>("op_99"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_7_cast_fp16 = reshape(shape = var_99, x = linear_2_cast_fp16)[name = tensor<string, []>("x_7_cast_fp16")];
+            tensor<int32, [4]> var_101 = const()[name = tensor<string, []>("op_101"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> var_105 = const()[name = tensor<string, []>("op_105"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_11_cast_fp16 = reshape(shape = var_105, x = linear_0_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
+            tensor<bool, []> attention_scores_1_transpose_x_0 = const()[name = tensor<string, []>("attention_scores_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attention_scores_1_transpose_y_0 = const()[name = tensor<string, []>("attention_scores_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_12_perm_0 = const()[name = tensor<string, []>("transpose_12_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_13_perm_0 = const()[name = tensor<string, []>("transpose_13_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 512]> transpose_13 = transpose(perm = transpose_13_perm_0, x = x_3_cast_fp16)[name = tensor<string, []>("transpose_33")];
+            tensor<fp16, [1, 8, 512, 64]> transpose_12 = transpose(perm = transpose_12_perm_0, x = x_11_cast_fp16)[name = tensor<string, []>("transpose_34")];
+            tensor<fp16, [1, 8, 512, 512]> attention_scores_1_cast_fp16 = matmul(transpose_x = attention_scores_1_transpose_x_0, transpose_y = attention_scores_1_transpose_y_0, x = transpose_12, y = transpose_13)[name = tensor<string, []>("attention_scores_1_cast_fp16")];
+            tensor<fp16, []> _inversed_attention_scores_3_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_attention_scores_3_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 512, 512]> _inversed_attention_scores_3_cast_fp16 = mul(x = attention_scores_1_cast_fp16, y = _inversed_attention_scores_3_y_0_to_fp16)[name = tensor<string, []>("_inversed_attention_scores_3_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> attention_scores_5_cast_fp16 = add(x = _inversed_attention_scores_3_cast_fp16, y = attention_mask_cast_fp16)[name = tensor<string, []>("attention_scores_5_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> bias_3_to_fp16 = const()[name = tensor<string, []>("bias_3_to_fp16"), val = tensor<fp16, [1, 8, 512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33363648)))];
+            tensor<fp16, [1, 8, 512, 512]> input_7_cast_fp16 = add(x = attention_scores_5_cast_fp16, y = bias_3_to_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> input_9_cast_fp16 = softmax(axis = var_16, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
+            tensor<bool, []> context_layer_1_transpose_x_0 = const()[name = tensor<string, []>("context_layer_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> context_layer_1_transpose_y_0 = const()[name = tensor<string, []>("context_layer_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 512, 64]> value_layer_1_cast_fp16 = transpose(perm = var_101, x = x_7_cast_fp16)[name = tensor<string, []>("transpose_35")];
+            tensor<fp16, [1, 8, 512, 64]> context_layer_1_cast_fp16 = matmul(transpose_x = context_layer_1_transpose_x_0, transpose_y = context_layer_1_transpose_y_0, x = input_9_cast_fp16, y = value_layer_1_cast_fp16)[name = tensor<string, []>("context_layer_1_cast_fp16")];
+            tensor<int32, [4]> var_118 = const()[name = tensor<string, []>("op_118"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_123 = const()[name = tensor<string, []>("op_123"), val = tensor<int32, [3]>([1, 512, 512])];
+            tensor<fp16, [1, 512, 8, 64]> var_119_cast_fp16 = transpose(perm = var_118, x = context_layer_1_cast_fp16)[name = tensor<string, []>("transpose_32")];
+            tensor<fp16, [1, 512, 512]> input_11_cast_fp16 = reshape(shape = var_123, x = var_119_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_0_attention_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37558016)))];
+            tensor<fp16, [512]> m_encoder_layer_0_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38082368)))];
+            tensor<fp16, [1, 512, 512]> linear_3_cast_fp16 = linear(bias = m_encoder_layer_0_attention_output_dense_bias_to_fp16, weight = m_encoder_layer_0_attention_output_dense_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
+            tensor<fp16, [1, 512, 512]> input_15_cast_fp16 = add(x = linear_3_cast_fp16, y = input_5_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
+            tensor<int32, [1]> input_17_axes_0 = const()[name = tensor<string, []>("input_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38083456)))];
+            tensor<fp16, [512]> m_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38084544)))];
+            tensor<fp16, [1, 512, 512]> input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = m_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
+            tensor<fp16, [4096, 512]> m_encoder_layer_0_mlp_gated_layers_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_mlp_gated_layers_weight_to_fp16"), val = tensor<fp16, [4096, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38085632)))];
+            tensor<fp16, [4096]> linear_4_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_4_bias_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42280000)))];
+            tensor<fp16, [1, 512, 4096]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = m_encoder_layer_0_mlp_gated_layers_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
+            tensor<int32, [3]> input_19_begin_0 = const()[name = tensor<string, []>("input_19_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> input_19_end_0 = const()[name = tensor<string, []>("input_19_end_0"), val = tensor<int32, [3]>([1, 512, 2048])];
+            tensor<bool, [3]> input_19_end_mask_0 = const()[name = tensor<string, []>("input_19_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 2048]> input_19_cast_fp16 = slice_by_index(begin = input_19_begin_0, end = input_19_end_0, end_mask = input_19_end_mask_0, x = linear_4_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
+            tensor<int32, [3]> non_gated_1_begin_0 = const()[name = tensor<string, []>("non_gated_1_begin_0"), val = tensor<int32, [3]>([0, 0, 2048])];
+            tensor<int32, [3]> non_gated_1_end_0 = const()[name = tensor<string, []>("non_gated_1_end_0"), val = tensor<int32, [3]>([1, 512, 4096])];
+            tensor<bool, [3]> non_gated_1_end_mask_0 = const()[name = tensor<string, []>("non_gated_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 2048]> non_gated_1_cast_fp16 = slice_by_index(begin = non_gated_1_begin_0, end = non_gated_1_end_0, end_mask = non_gated_1_end_mask_0, x = linear_4_cast_fp16)[name = tensor<string, []>("non_gated_1_cast_fp16")];
+            tensor<string, []> var_147_mode_0 = const()[name = tensor<string, []>("op_147_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 512, 2048]> var_147_cast_fp16 = gelu(mode = var_147_mode_0, x = input_19_cast_fp16)[name = tensor<string, []>("op_147_cast_fp16")];
+            tensor<fp16, [1, 512, 2048]> input_21_cast_fp16 = mul(x = var_147_cast_fp16, y = non_gated_1_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
+            tensor<fp16, [512, 2048]> m_encoder_layer_0_mlp_wo_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_mlp_wo_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42288256)))];
+            tensor<fp16, [512]> m_encoder_layer_0_mlp_wo_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_mlp_wo_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44385472)))];
+            tensor<fp16, [1, 512, 512]> linear_5_cast_fp16 = linear(bias = m_encoder_layer_0_mlp_wo_bias_to_fp16, weight = m_encoder_layer_0_mlp_wo_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
+            tensor<fp16, [1, 512, 512]> input_25_cast_fp16 = add(x = linear_5_cast_fp16, y = input_17_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
+            tensor<int32, [1]> input_27_axes_0 = const()[name = tensor<string, []>("input_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_0_mlp_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_mlp_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44386560)))];
+            tensor<fp16, [512]> m_encoder_layer_0_mlp_layernorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_mlp_layernorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44387648)))];
+            tensor<fp16, [1, 512, 512]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = m_encoder_layer_0_mlp_layernorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_0_mlp_layernorm_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_1_attention_self_query_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_query_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44388736)))];
+            tensor<fp16, [512]> m_encoder_layer_1_attention_self_query_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_query_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44913088)))];
+            tensor<fp16, [1, 512, 512]> linear_6_cast_fp16 = linear(bias = m_encoder_layer_1_attention_self_query_bias_to_fp16, weight = m_encoder_layer_1_attention_self_query_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_1_attention_self_key_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_key_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44914176)))];
+            tensor<fp16, [512]> m_encoder_layer_1_attention_self_key_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_key_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45438528)))];
+            tensor<fp16, [1, 512, 512]> linear_7_cast_fp16 = linear(bias = m_encoder_layer_1_attention_self_key_bias_to_fp16, weight = m_encoder_layer_1_attention_self_key_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
+            tensor<int32, [4]> var_173 = const()[name = tensor<string, []>("op_173"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_15_cast_fp16 = reshape(shape = var_173, x = linear_7_cast_fp16)[name = tensor<string, []>("x_15_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_1_attention_self_value_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_value_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45439616)))];
+            tensor<fp16, [512]> m_encoder_layer_1_attention_self_value_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_value_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45963968)))];
+            tensor<fp16, [1, 512, 512]> linear_8_cast_fp16 = linear(bias = m_encoder_layer_1_attention_self_value_bias_to_fp16, weight = m_encoder_layer_1_attention_self_value_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
+            tensor<int32, [4]> var_182 = const()[name = tensor<string, []>("op_182"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_19_cast_fp16 = reshape(shape = var_182, x = linear_8_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")];
+            tensor<int32, [4]> var_184 = const()[name = tensor<string, []>("op_184"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> var_188 = const()[name = tensor<string, []>("op_188"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_23_cast_fp16 = reshape(shape = var_188, x = linear_6_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")];
+            tensor<bool, []> attention_scores_7_transpose_x_0 = const()[name = tensor<string, []>("attention_scores_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attention_scores_7_transpose_y_0 = const()[name = tensor<string, []>("attention_scores_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_14_perm_0 = const()[name = tensor<string, []>("transpose_14_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_15_perm_0 = const()[name = tensor<string, []>("transpose_15_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 512]> transpose_15 = transpose(perm = transpose_15_perm_0, x = x_15_cast_fp16)[name = tensor<string, []>("transpose_29")];
+            tensor<fp16, [1, 8, 512, 64]> transpose_14 = transpose(perm = transpose_14_perm_0, x = x_23_cast_fp16)[name = tensor<string, []>("transpose_30")];
+            tensor<fp16, [1, 8, 512, 512]> attention_scores_7_cast_fp16 = matmul(transpose_x = attention_scores_7_transpose_x_0, transpose_y = attention_scores_7_transpose_y_0, x = transpose_14, y = transpose_15)[name = tensor<string, []>("attention_scores_7_cast_fp16")];
+            tensor<fp16, []> _inversed_attention_scores_9_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_attention_scores_9_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 512, 512]> _inversed_attention_scores_9_cast_fp16 = mul(x = attention_scores_7_cast_fp16, y = _inversed_attention_scores_9_y_0_to_fp16)[name = tensor<string, []>("_inversed_attention_scores_9_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> attention_scores_11_cast_fp16 = add(x = _inversed_attention_scores_9_cast_fp16, y = attention_mask_cast_fp16)[name = tensor<string, []>("attention_scores_11_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> input_29_cast_fp16 = add(x = attention_scores_11_cast_fp16, y = bias_3_to_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> input_31_cast_fp16 = softmax(axis = var_16, x = input_29_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
+            tensor<bool, []> context_layer_5_transpose_x_0 = const()[name = tensor<string, []>("context_layer_5_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> context_layer_5_transpose_y_0 = const()[name = tensor<string, []>("context_layer_5_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 512, 64]> value_layer_3_cast_fp16 = transpose(perm = var_184, x = x_19_cast_fp16)[name = tensor<string, []>("transpose_31")];
+            tensor<fp16, [1, 8, 512, 64]> context_layer_5_cast_fp16 = matmul(transpose_x = context_layer_5_transpose_x_0, transpose_y = context_layer_5_transpose_y_0, x = input_31_cast_fp16, y = value_layer_3_cast_fp16)[name = tensor<string, []>("context_layer_5_cast_fp16")];
+            tensor<int32, [4]> var_201 = const()[name = tensor<string, []>("op_201"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_206 = const()[name = tensor<string, []>("op_206"), val = tensor<int32, [3]>([1, 512, 512])];
+            tensor<fp16, [1, 512, 8, 64]> var_202_cast_fp16 = transpose(perm = var_201, x = context_layer_5_cast_fp16)[name = tensor<string, []>("transpose_28")];
+            tensor<fp16, [1, 512, 512]> input_33_cast_fp16 = reshape(shape = var_206, x = var_202_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_1_attention_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(45965056)))];
+            tensor<fp16, [512]> m_encoder_layer_1_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46489408)))];
+            tensor<fp16, [1, 512, 512]> linear_9_cast_fp16 = linear(bias = m_encoder_layer_1_attention_output_dense_bias_to_fp16, weight = m_encoder_layer_1_attention_output_dense_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
+            tensor<fp16, [1, 512, 512]> input_37_cast_fp16 = add(x = linear_9_cast_fp16, y = input_27_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
+            tensor<int32, [1]> input_39_axes_0 = const()[name = tensor<string, []>("input_39_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46490496)))];
+            tensor<fp16, [512]> m_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46491584)))];
+            tensor<fp16, [1, 512, 512]> input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = m_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
+            tensor<fp16, [4096, 512]> m_encoder_layer_1_mlp_gated_layers_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_mlp_gated_layers_weight_to_fp16"), val = tensor<fp16, [4096, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46492672)))];
+            tensor<fp16, [1, 512, 4096]> linear_10_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = m_encoder_layer_1_mlp_gated_layers_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
+            tensor<int32, [3]> input_41_begin_0 = const()[name = tensor<string, []>("input_41_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> input_41_end_0 = const()[name = tensor<string, []>("input_41_end_0"), val = tensor<int32, [3]>([1, 512, 2048])];
+            tensor<bool, [3]> input_41_end_mask_0 = const()[name = tensor<string, []>("input_41_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 2048]> input_41_cast_fp16 = slice_by_index(begin = input_41_begin_0, end = input_41_end_0, end_mask = input_41_end_mask_0, x = linear_10_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
+            tensor<int32, [3]> non_gated_3_begin_0 = const()[name = tensor<string, []>("non_gated_3_begin_0"), val = tensor<int32, [3]>([0, 0, 2048])];
+            tensor<int32, [3]> non_gated_3_end_0 = const()[name = tensor<string, []>("non_gated_3_end_0"), val = tensor<int32, [3]>([1, 512, 4096])];
+            tensor<bool, [3]> non_gated_3_end_mask_0 = const()[name = tensor<string, []>("non_gated_3_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 2048]> non_gated_3_cast_fp16 = slice_by_index(begin = non_gated_3_begin_0, end = non_gated_3_end_0, end_mask = non_gated_3_end_mask_0, x = linear_10_cast_fp16)[name = tensor<string, []>("non_gated_3_cast_fp16")];
+            tensor<string, []> var_230_mode_0 = const()[name = tensor<string, []>("op_230_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 512, 2048]> var_230_cast_fp16 = gelu(mode = var_230_mode_0, x = input_41_cast_fp16)[name = tensor<string, []>("op_230_cast_fp16")];
+            tensor<fp16, [1, 512, 2048]> input_43_cast_fp16 = mul(x = var_230_cast_fp16, y = non_gated_3_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
+            tensor<fp16, [512, 2048]> m_encoder_layer_1_mlp_wo_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_mlp_wo_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50687040)))];
+            tensor<fp16, [512]> m_encoder_layer_1_mlp_wo_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_mlp_wo_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52784256)))];
+            tensor<fp16, [1, 512, 512]> linear_11_cast_fp16 = linear(bias = m_encoder_layer_1_mlp_wo_bias_to_fp16, weight = m_encoder_layer_1_mlp_wo_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
+            tensor<fp16, [1, 512, 512]> input_47_cast_fp16 = add(x = linear_11_cast_fp16, y = input_39_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
+            tensor<int32, [1]> input_49_axes_0 = const()[name = tensor<string, []>("input_49_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_1_mlp_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_mlp_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52785344)))];
+            tensor<fp16, [512]> m_encoder_layer_1_mlp_layernorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_mlp_layernorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52786432)))];
+            tensor<fp16, [1, 512, 512]> input_49_cast_fp16 = layer_norm(axes = input_49_axes_0, beta = m_encoder_layer_1_mlp_layernorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_1_mlp_layernorm_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_2_attention_self_query_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_query_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52787520)))];
+            tensor<fp16, [512]> m_encoder_layer_2_attention_self_query_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_query_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53311872)))];
+            tensor<fp16, [1, 512, 512]> linear_12_cast_fp16 = linear(bias = m_encoder_layer_2_attention_self_query_bias_to_fp16, weight = m_encoder_layer_2_attention_self_query_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_2_attention_self_key_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_key_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53312960)))];
+            tensor<fp16, [512]> m_encoder_layer_2_attention_self_key_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_key_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53837312)))];
+            tensor<fp16, [1, 512, 512]> linear_13_cast_fp16 = linear(bias = m_encoder_layer_2_attention_self_key_bias_to_fp16, weight = m_encoder_layer_2_attention_self_key_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
+            tensor<int32, [4]> var_256 = const()[name = tensor<string, []>("op_256"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_27_cast_fp16 = reshape(shape = var_256, x = linear_13_cast_fp16)[name = tensor<string, []>("x_27_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_2_attention_self_value_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_value_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53838400)))];
+            tensor<fp16, [512]> m_encoder_layer_2_attention_self_value_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_value_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54362752)))];
+            tensor<fp16, [1, 512, 512]> linear_14_cast_fp16 = linear(bias = m_encoder_layer_2_attention_self_value_bias_to_fp16, weight = m_encoder_layer_2_attention_self_value_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
+            tensor<int32, [4]> var_265 = const()[name = tensor<string, []>("op_265"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_31_cast_fp16 = reshape(shape = var_265, x = linear_14_cast_fp16)[name = tensor<string, []>("x_31_cast_fp16")];
+            tensor<int32, [4]> var_267 = const()[name = tensor<string, []>("op_267"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> var_271 = const()[name = tensor<string, []>("op_271"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_35_cast_fp16 = reshape(shape = var_271, x = linear_12_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")];
+            tensor<bool, []> attention_scores_13_transpose_x_0 = const()[name = tensor<string, []>("attention_scores_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attention_scores_13_transpose_y_0 = const()[name = tensor<string, []>("attention_scores_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_16_perm_0 = const()[name = tensor<string, []>("transpose_16_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_17_perm_0 = const()[name = tensor<string, []>("transpose_17_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 512]> transpose_17 = transpose(perm = transpose_17_perm_0, x = x_27_cast_fp16)[name = tensor<string, []>("transpose_25")];
+            tensor<fp16, [1, 8, 512, 64]> transpose_16 = transpose(perm = transpose_16_perm_0, x = x_35_cast_fp16)[name = tensor<string, []>("transpose_26")];
+            tensor<fp16, [1, 8, 512, 512]> attention_scores_13_cast_fp16 = matmul(transpose_x = attention_scores_13_transpose_x_0, transpose_y = attention_scores_13_transpose_y_0, x = transpose_16, y = transpose_17)[name = tensor<string, []>("attention_scores_13_cast_fp16")];
+            tensor<fp16, []> _inversed_attention_scores_15_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_attention_scores_15_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 512, 512]> _inversed_attention_scores_15_cast_fp16 = mul(x = attention_scores_13_cast_fp16, y = _inversed_attention_scores_15_y_0_to_fp16)[name = tensor<string, []>("_inversed_attention_scores_15_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> attention_scores_17_cast_fp16 = add(x = _inversed_attention_scores_15_cast_fp16, y = attention_mask_cast_fp16)[name = tensor<string, []>("attention_scores_17_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> input_51_cast_fp16 = add(x = attention_scores_17_cast_fp16, y = bias_3_to_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> input_53_cast_fp16 = softmax(axis = var_16, x = input_51_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
+            tensor<bool, []> context_layer_9_transpose_x_0 = const()[name = tensor<string, []>("context_layer_9_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> context_layer_9_transpose_y_0 = const()[name = tensor<string, []>("context_layer_9_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 512, 64]> value_layer_5_cast_fp16 = transpose(perm = var_267, x = x_31_cast_fp16)[name = tensor<string, []>("transpose_27")];
+            tensor<fp16, [1, 8, 512, 64]> context_layer_9_cast_fp16 = matmul(transpose_x = context_layer_9_transpose_x_0, transpose_y = context_layer_9_transpose_y_0, x = input_53_cast_fp16, y = value_layer_5_cast_fp16)[name = tensor<string, []>("context_layer_9_cast_fp16")];
+            tensor<int32, [4]> var_284 = const()[name = tensor<string, []>("op_284"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_289 = const()[name = tensor<string, []>("op_289"), val = tensor<int32, [3]>([1, 512, 512])];
+            tensor<fp16, [1, 512, 8, 64]> var_285_cast_fp16 = transpose(perm = var_284, x = context_layer_9_cast_fp16)[name = tensor<string, []>("transpose_24")];
+            tensor<fp16, [1, 512, 512]> input_55_cast_fp16 = reshape(shape = var_289, x = var_285_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_2_attention_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54363840)))];
+            tensor<fp16, [512]> m_encoder_layer_2_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54888192)))];
+            tensor<fp16, [1, 512, 512]> linear_15_cast_fp16 = linear(bias = m_encoder_layer_2_attention_output_dense_bias_to_fp16, weight = m_encoder_layer_2_attention_output_dense_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("linear_15_cast_fp16")];
+            tensor<fp16, [1, 512, 512]> input_59_cast_fp16 = add(x = linear_15_cast_fp16, y = input_49_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
+            tensor<int32, [1]> input_61_axes_0 = const()[name = tensor<string, []>("input_61_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54889280)))];
+            tensor<fp16, [512]> m_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54890368)))];
+            tensor<fp16, [1, 512, 512]> input_61_cast_fp16 = layer_norm(axes = input_61_axes_0, beta = m_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
+            tensor<fp16, [4096, 512]> m_encoder_layer_2_mlp_gated_layers_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_mlp_gated_layers_weight_to_fp16"), val = tensor<fp16, [4096, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54891456)))];
+            tensor<fp16, [1, 512, 4096]> linear_16_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = m_encoder_layer_2_mlp_gated_layers_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
+            tensor<int32, [3]> input_63_begin_0 = const()[name = tensor<string, []>("input_63_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> input_63_end_0 = const()[name = tensor<string, []>("input_63_end_0"), val = tensor<int32, [3]>([1, 512, 2048])];
+            tensor<bool, [3]> input_63_end_mask_0 = const()[name = tensor<string, []>("input_63_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 2048]> input_63_cast_fp16 = slice_by_index(begin = input_63_begin_0, end = input_63_end_0, end_mask = input_63_end_mask_0, x = linear_16_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
+            tensor<int32, [3]> non_gated_5_begin_0 = const()[name = tensor<string, []>("non_gated_5_begin_0"), val = tensor<int32, [3]>([0, 0, 2048])];
+            tensor<int32, [3]> non_gated_5_end_0 = const()[name = tensor<string, []>("non_gated_5_end_0"), val = tensor<int32, [3]>([1, 512, 4096])];
+            tensor<bool, [3]> non_gated_5_end_mask_0 = const()[name = tensor<string, []>("non_gated_5_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 2048]> non_gated_5_cast_fp16 = slice_by_index(begin = non_gated_5_begin_0, end = non_gated_5_end_0, end_mask = non_gated_5_end_mask_0, x = linear_16_cast_fp16)[name = tensor<string, []>("non_gated_5_cast_fp16")];
+            tensor<string, []> var_313_mode_0 = const()[name = tensor<string, []>("op_313_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 512, 2048]> var_313_cast_fp16 = gelu(mode = var_313_mode_0, x = input_63_cast_fp16)[name = tensor<string, []>("op_313_cast_fp16")];
+            tensor<fp16, [1, 512, 2048]> input_65_cast_fp16 = mul(x = var_313_cast_fp16, y = non_gated_5_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
+            tensor<fp16, [512, 2048]> m_encoder_layer_2_mlp_wo_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_mlp_wo_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59085824)))];
+            tensor<fp16, [512]> m_encoder_layer_2_mlp_wo_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_mlp_wo_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61183040)))];
+            tensor<fp16, [1, 512, 512]> linear_17_cast_fp16 = linear(bias = m_encoder_layer_2_mlp_wo_bias_to_fp16, weight = m_encoder_layer_2_mlp_wo_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
+            tensor<fp16, [1, 512, 512]> input_69_cast_fp16 = add(x = linear_17_cast_fp16, y = input_61_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
+            tensor<int32, [1]> input_71_axes_0 = const()[name = tensor<string, []>("input_71_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_2_mlp_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_mlp_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61184128)))];
+            tensor<fp16, [512]> m_encoder_layer_2_mlp_layernorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_mlp_layernorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61185216)))];
+            tensor<fp16, [1, 512, 512]> input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = m_encoder_layer_2_mlp_layernorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_2_mlp_layernorm_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_3_attention_self_query_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_query_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61186304)))];
+            tensor<fp16, [512]> m_encoder_layer_3_attention_self_query_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_query_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61710656)))];
+            tensor<fp16, [1, 512, 512]> linear_18_cast_fp16 = linear(bias = m_encoder_layer_3_attention_self_query_bias_to_fp16, weight = m_encoder_layer_3_attention_self_query_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_3_attention_self_key_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_key_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61711744)))];
+            tensor<fp16, [512]> m_encoder_layer_3_attention_self_key_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_key_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62236096)))];
+            tensor<fp16, [1, 512, 512]> linear_19_cast_fp16 = linear(bias = m_encoder_layer_3_attention_self_key_bias_to_fp16, weight = m_encoder_layer_3_attention_self_key_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
+            tensor<int32, [4]> var_339 = const()[name = tensor<string, []>("op_339"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_39_cast_fp16 = reshape(shape = var_339, x = linear_19_cast_fp16)[name = tensor<string, []>("x_39_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_3_attention_self_value_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_value_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62237184)))];
+            tensor<fp16, [512]> m_encoder_layer_3_attention_self_value_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_value_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62761536)))];
+            tensor<fp16, [1, 512, 512]> linear_20_cast_fp16 = linear(bias = m_encoder_layer_3_attention_self_value_bias_to_fp16, weight = m_encoder_layer_3_attention_self_value_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
+            tensor<int32, [4]> var_348 = const()[name = tensor<string, []>("op_348"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_43_cast_fp16 = reshape(shape = var_348, x = linear_20_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")];
+            tensor<int32, [4]> var_350 = const()[name = tensor<string, []>("op_350"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> var_354 = const()[name = tensor<string, []>("op_354"), val = tensor<int32, [4]>([1, 512, 8, 64])];
+            tensor<fp16, [1, 512, 8, 64]> x_cast_fp16 = reshape(shape = var_354, x = linear_18_cast_fp16)[name = tensor<string, []>("x_cast_fp16")];
+            tensor<bool, []> attention_scores_19_transpose_x_0 = const()[name = tensor<string, []>("attention_scores_19_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attention_scores_19_transpose_y_0 = const()[name = tensor<string, []>("attention_scores_19_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_18_perm_0 = const()[name = tensor<string, []>("transpose_18_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_19_perm_0 = const()[name = tensor<string, []>("transpose_19_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 512]> transpose_19 = transpose(perm = transpose_19_perm_0, x = x_39_cast_fp16)[name = tensor<string, []>("transpose_21")];
+            tensor<fp16, [1, 8, 512, 64]> transpose_18 = transpose(perm = transpose_18_perm_0, x = x_cast_fp16)[name = tensor<string, []>("transpose_22")];
+            tensor<fp16, [1, 8, 512, 512]> attention_scores_19_cast_fp16 = matmul(transpose_x = attention_scores_19_transpose_x_0, transpose_y = attention_scores_19_transpose_y_0, x = transpose_18, y = transpose_19)[name = tensor<string, []>("attention_scores_19_cast_fp16")];
+            tensor<fp16, []> _inversed_attention_scores_21_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_attention_scores_21_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 512, 512]> _inversed_attention_scores_21_cast_fp16 = mul(x = attention_scores_19_cast_fp16, y = _inversed_attention_scores_21_y_0_to_fp16)[name = tensor<string, []>("_inversed_attention_scores_21_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> attention_scores_cast_fp16 = add(x = _inversed_attention_scores_21_cast_fp16, y = attention_mask_cast_fp16)[name = tensor<string, []>("attention_scores_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> input_73_cast_fp16 = add(x = attention_scores_cast_fp16, y = bias_3_to_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
+            tensor<fp16, [1, 8, 512, 512]> input_75_cast_fp16 = softmax(axis = var_16, x = input_73_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
+            tensor<bool, []> context_layer_13_transpose_x_0 = const()[name = tensor<string, []>("context_layer_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> context_layer_13_transpose_y_0 = const()[name = tensor<string, []>("context_layer_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 512, 64]> value_layer_cast_fp16 = transpose(perm = var_350, x = x_43_cast_fp16)[name = tensor<string, []>("transpose_23")];
+            tensor<fp16, [1, 8, 512, 64]> context_layer_13_cast_fp16 = matmul(transpose_x = context_layer_13_transpose_x_0, transpose_y = context_layer_13_transpose_y_0, x = input_75_cast_fp16, y = value_layer_cast_fp16)[name = tensor<string, []>("context_layer_13_cast_fp16")];
+            tensor<int32, [4]> var_367 = const()[name = tensor<string, []>("op_367"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_372 = const()[name = tensor<string, []>("op_372"), val = tensor<int32, [3]>([1, 512, 512])];
+            tensor<fp16, [1, 512, 8, 64]> var_368_cast_fp16 = transpose(perm = var_367, x = context_layer_13_cast_fp16)[name = tensor<string, []>("transpose_20")];
+            tensor<fp16, [1, 512, 512]> input_77_cast_fp16 = reshape(shape = var_372, x = var_368_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_3_attention_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62762624)))];
+            tensor<fp16, [512]> m_encoder_layer_3_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63286976)))];
+            tensor<fp16, [1, 512, 512]> linear_21_cast_fp16 = linear(bias = m_encoder_layer_3_attention_output_dense_bias_to_fp16, weight = m_encoder_layer_3_attention_output_dense_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
+            tensor<fp16, [1, 512, 512]> input_81_cast_fp16 = add(x = linear_21_cast_fp16, y = input_71_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
+            tensor<int32, [1]> input_83_axes_0 = const()[name = tensor<string, []>("input_83_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63288064)))];
+            tensor<fp16, [512]> m_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63289152)))];
+            tensor<fp16, [1, 512, 512]> input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = m_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
+            tensor<fp16, [4096, 512]> m_encoder_layer_3_mlp_gated_layers_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_mlp_gated_layers_weight_to_fp16"), val = tensor<fp16, [4096, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63290240)))];
+            tensor<fp16, [1, 512, 4096]> linear_22_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = m_encoder_layer_3_mlp_gated_layers_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
+            tensor<int32, [3]> input_85_begin_0 = const()[name = tensor<string, []>("input_85_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> input_85_end_0 = const()[name = tensor<string, []>("input_85_end_0"), val = tensor<int32, [3]>([1, 512, 2048])];
+            tensor<bool, [3]> input_85_end_mask_0 = const()[name = tensor<string, []>("input_85_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 512, 2048]> input_85_cast_fp16 = slice_by_index(begin = input_85_begin_0, end = input_85_end_0, end_mask = input_85_end_mask_0, x = linear_22_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
+            tensor<int32, [3]> non_gated_begin_0 = const()[name = tensor<string, []>("non_gated_begin_0"), val = tensor<int32, [3]>([0, 0, 2048])];
+            tensor<int32, [3]> non_gated_end_0 = const()[name = tensor<string, []>("non_gated_end_0"), val = tensor<int32, [3]>([1, 512, 4096])];
+            tensor<bool, [3]> non_gated_end_mask_0 = const()[name = tensor<string, []>("non_gated_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 512, 2048]> non_gated_cast_fp16 = slice_by_index(begin = non_gated_begin_0, end = non_gated_end_0, end_mask = non_gated_end_mask_0, x = linear_22_cast_fp16)[name = tensor<string, []>("non_gated_cast_fp16")];
+            tensor<string, []> var_396_mode_0 = const()[name = tensor<string, []>("op_396_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 512, 2048]> var_396_cast_fp16 = gelu(mode = var_396_mode_0, x = input_85_cast_fp16)[name = tensor<string, []>("op_396_cast_fp16")];
+            tensor<fp16, [1, 512, 2048]> input_87_cast_fp16 = mul(x = var_396_cast_fp16, y = non_gated_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
+            tensor<fp16, [512, 2048]> m_encoder_layer_3_mlp_wo_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_mlp_wo_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67484608)))];
+            tensor<fp16, [512]> m_encoder_layer_3_mlp_wo_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_mlp_wo_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69581824)))];
+            tensor<fp16, [1, 512, 512]> linear_23_cast_fp16 = linear(bias = m_encoder_layer_3_mlp_wo_bias_to_fp16, weight = m_encoder_layer_3_mlp_wo_weight_to_fp16, x = input_87_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
+            tensor<fp16, [1, 512, 512]> input_91_cast_fp16 = add(x = linear_23_cast_fp16, y = input_83_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
+            tensor<int32, [1]> hidden_states_axes_0 = const()[name = tensor<string, []>("hidden_states_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_3_mlp_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_mlp_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69582912)))];
+            tensor<fp16, [512]> m_encoder_layer_3_mlp_layernorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_mlp_layernorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69584000)))];
+            tensor<fp16, [1, 512, 512]> hidden_states_cast_fp16 = layer_norm(axes = hidden_states_axes_0, beta = m_encoder_layer_3_mlp_layernorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_3_mlp_layernorm_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")];
+            tensor<int32, [1]> var_408_axes_0 = const()[name = tensor<string, []>("op_408_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 512, 1]> var_408 = expand_dims(axes = var_408_axes_0, x = cast_0)[name = tensor<string, []>("op_408")];
+            tensor<string, []> cast_28_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_28_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 512, 1]> var_408_to_fp16 = cast(dtype = cast_28_to_fp16_dtype_0, x = var_408)[name = tensor<string, []>("cast_34")];
+            tensor<fp16, [1, 512, 512]> var_414_cast_fp16 = mul(x = hidden_states_cast_fp16, y = var_408_to_fp16)[name = tensor<string, []>("op_414_cast_fp16")];
+            tensor<int32, [1]> var_419_axes_0 = const()[name = tensor<string, []>("op_419_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, []> var_419_keep_dims_0 = const()[name = tensor<string, []>("op_419_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512]> var_419_cast_fp16 = reduce_sum(axes = var_419_axes_0, keep_dims = var_419_keep_dims_0, x = var_414_cast_fp16)[name = tensor<string, []>("op_419_cast_fp16")];
+            tensor<int32, [1]> var_424_axes_0 = const()[name = tensor<string, []>("op_424_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, []> var_424_keep_dims_0 = const()[name = tensor<string, []>("op_424_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1]> var_424_cast_fp16 = reduce_sum(axes = var_424_axes_0, keep_dims = var_424_keep_dims_0, x = var_408_to_fp16)[name = tensor<string, []>("op_424_cast_fp16")];
+            tensor<fp16, []> var_425_to_fp16 = const()[name = tensor<string, []>("op_425_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, []> const_35_to_fp16 = const()[name = tensor<string, []>("const_35_to_fp16"), val = tensor<fp16, []>(inf)];
+            tensor<fp16, [1, 1]> clip_0_cast_fp16 = clip(alpha = var_425_to_fp16, beta = const_35_to_fp16, x = var_424_cast_fp16)[name = tensor<string, []>("clip_0_cast_fp16")];
+            tensor<fp16, [1, 512]> input_cast_fp16 = real_div(x = var_419_cast_fp16, y = clip_0_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [1]> var_431 = const()[name = tensor<string, []>("op_431"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> var_432 = const()[name = tensor<string, []>("op_432"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 1]> var_434_cast_fp16 = reduce_l2_norm(axes = var_431, keep_dims = var_432, x = input_cast_fp16)[name = tensor<string, []>("op_434_cast_fp16")];
+            tensor<fp16, []> var_435_to_fp16 = const()[name = tensor<string, []>("op_435_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [1, 1]> var_436_cast_fp16 = maximum(x = var_434_cast_fp16, y = var_435_to_fp16)[name = tensor<string, []>("op_436_cast_fp16")];
+            tensor<int32, [2]> denom_reps_0 = const()[name = tensor<string, []>("denom_reps_0"), val = tensor<int32, [2]>([1, 512])];
+            tensor<fp16, [1, 512]> denom_cast_fp16 = tile(reps = denom_reps_0, x = var_436_cast_fp16)[name = tensor<string, []>("denom_cast_fp16")];
+            tensor<fp16, [1, 512]> embedding = real_div(x = input_cast_fp16, y = denom_cast_fp16)[name = tensor<string, []>("op_438_cast_fp16")];
+        } -> (embedding);
+}

jina_embedding.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40bc6b82fb23ad0c7651c0129ebb53309cf7f8e85d14e81752d517572790945b
+size 69585088

jina_embedding_128.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0161a3aacffe1a62c443927a83a1f337f1b9f75aed0f71ef561bfcfc1abc03fc
+size 243

jina_embedding_128.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2aea8d0e1ba9eeab169a290e965c9c19ba9ae6b4da42153f5becd7edc3cc2799
+size 318

jina_embedding_128.mlmodelc/metadata.json

@@ -0,0 +1,80 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 512)",
+        "shortDescription" : "",
+        "shape" : "[1, 512]",
+        "name" : "embedding",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.mul" : 10,
+      "Ios17.notEqual" : 1,
+      "Ios17.reshape" : 16,
+      "Ios16.softmax" : 4,
+      "Ios17.matmul" : 8,
+      "Ios17.maximum" : 1,
+      "Ios17.transpose" : 16,
+      "Ios17.expandDims" : 3,
+      "Ios17.add" : 17,
+      "Tile" : 1,
+      "Ios17.sliceByIndex" : 8,
+      "Ios16.reduceSum" : 2,
+      "Ios17.gather" : 1,
+      "Ios17.layerNorm" : 9,
+      "Ios17.clip" : 1,
+      "Ios17.sub" : 1,
+      "Ios16.gelu" : 4,
+      "Ios17.cast" : 4,
+      "Ios17.realDiv" : 2,
+      "Ios17.linear" : 24,
+      "Ios16.reduceL2Norm" : 1
+    },
+    "computePrecision" : "Mixed (Float16, Int16, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.5.1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "input_ids",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "jina_embedding_ane_128",
+    "method" : "predict"
+  }
+]

jina_embedding_128.mlmodelc/model.mil

@@ -0,0 +1,320 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<int32, [1, 128]> input_ids) {
+            tensor<int32, []> var_3 = const()[name = tensor<string, []>("op_3"), val = tensor<int32, []>(0)];
+            tensor<bool, [1, 128]> var_4 = not_equal(x = input_ids, y = var_3)[name = tensor<string, []>("op_4")];
+            tensor<string, []> cast_0_dtype_0 = const()[name = tensor<string, []>("cast_0_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, []> var_16 = const()[name = tensor<string, []>("op_16"), val = tensor<int32, []>(-1)];
+            tensor<int32, [1]> var_41_axes_0 = const()[name = tensor<string, []>("op_41_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 128]> cast_0 = cast(dtype = cast_0_dtype_0, x = var_4)[name = tensor<string, []>("cast_37")];
+            tensor<int32, [1, 1, 128]> var_41 = expand_dims(axes = var_41_axes_0, x = cast_0)[name = tensor<string, []>("op_41")];
+            tensor<int32, [1]> var_42_axes_0 = const()[name = tensor<string, []>("op_42_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<int32, [1, 1, 1, 128]> var_42 = expand_dims(axes = var_42_axes_0, x = var_41)[name = tensor<string, []>("op_42")];
+            tensor<fp16, []> var_21_to_fp16 = const()[name = tensor<string, []>("op_21_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<string, []> cast_1_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 1, 1, 128]> var_42_to_fp16 = cast(dtype = cast_1_to_fp16_dtype_0, x = var_42)[name = tensor<string, []>("cast_36")];
+            tensor<fp16, [1, 1, 1, 128]> var_45_cast_fp16 = sub(x = var_21_to_fp16, y = var_42_to_fp16)[name = tensor<string, []>("op_45_cast_fp16")];
+            tensor<fp16, []> var_46_to_fp16 = const()[name = tensor<string, []>("op_46_to_fp16"), val = tensor<fp16, []>(-0x1.d4cp+14)];
+            tensor<fp16, [1, 1, 1, 128]> attention_mask_cast_fp16 = mul(x = var_45_cast_fp16, y = var_46_to_fp16)[name = tensor<string, []>("attention_mask_cast_fp16")];
+            tensor<int32, []> inputs_embeds_axis_0 = const()[name = tensor<string, []>("inputs_embeds_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> inputs_embeds_batch_dims_0 = const()[name = tensor<string, []>("inputs_embeds_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> inputs_embeds_validate_indices_0 = const()[name = tensor<string, []>("inputs_embeds_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [30528, 512]> m_embeddings_word_embeddings_weight_to_fp16 = const()[name = tensor<string, []>("m_embeddings_word_embeddings_weight_to_fp16"), val = tensor<fp16, [30528, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<string, []> input_ids_to_int16_dtype_0 = const()[name = tensor<string, []>("input_ids_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 128]> input_ids_to_int16 = cast(dtype = input_ids_to_int16_dtype_0, x = input_ids)[name = tensor<string, []>("cast_35")];
+            tensor<fp16, [1, 128, 512]> inputs_embeds_cast_fp16_cast_uint16 = gather(axis = inputs_embeds_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = input_ids_to_int16, validate_indices = inputs_embeds_validate_indices_0, x = m_embeddings_word_embeddings_weight_to_fp16)[name = tensor<string, []>("inputs_embeds_cast_fp16_cast_uint16")];
+            tensor<fp16, [1, 128, 512]> token_type_embeddings_1_to_fp16 = const()[name = tensor<string, []>("token_type_embeddings_1_to_fp16"), val = tensor<fp16, [1, 128, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31260800)))];
+            tensor<fp16, [1, 128, 512]> input_3_cast_fp16 = add(x = inputs_embeds_cast_fp16_cast_uint16, y = token_type_embeddings_1_to_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<int32, [1]> input_5_axes_0 = const()[name = tensor<string, []>("input_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_embeddings_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_embeddings_LayerNorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31391936)))];
+            tensor<fp16, [512]> m_embeddings_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_embeddings_LayerNorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31393024)))];
+            tensor<fp16, []> var_18_to_fp16 = const()[name = tensor<string, []>("op_18_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [1, 128, 512]> input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = m_embeddings_LayerNorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_embeddings_LayerNorm_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_0_attention_self_query_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_query_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31394112)))];
+            tensor<fp16, [512]> m_encoder_layer_0_attention_self_query_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_query_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31918464)))];
+            tensor<fp16, [1, 128, 512]> linear_0_cast_fp16 = linear(bias = m_encoder_layer_0_attention_self_query_bias_to_fp16, weight = m_encoder_layer_0_attention_self_query_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_0_attention_self_key_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_key_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31919552)))];
+            tensor<fp16, [512]> m_encoder_layer_0_attention_self_key_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_key_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32443904)))];
+            tensor<fp16, [1, 128, 512]> linear_1_cast_fp16 = linear(bias = m_encoder_layer_0_attention_self_key_bias_to_fp16, weight = m_encoder_layer_0_attention_self_key_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<int32, [4]> var_90 = const()[name = tensor<string, []>("op_90"), val = tensor<int32, [4]>([1, 128, 8, 64])];
+            tensor<fp16, [1, 128, 8, 64]> x_3_cast_fp16 = reshape(shape = var_90, x = linear_1_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_0_attention_self_value_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_value_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32444992)))];
+            tensor<fp16, [512]> m_encoder_layer_0_attention_self_value_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_value_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32969344)))];
+            tensor<fp16, [1, 128, 512]> linear_2_cast_fp16 = linear(bias = m_encoder_layer_0_attention_self_value_bias_to_fp16, weight = m_encoder_layer_0_attention_self_value_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<int32, [4]> var_99 = const()[name = tensor<string, []>("op_99"), val = tensor<int32, [4]>([1, 128, 8, 64])];
+            tensor<fp16, [1, 128, 8, 64]> x_7_cast_fp16 = reshape(shape = var_99, x = linear_2_cast_fp16)[name = tensor<string, []>("x_7_cast_fp16")];
+            tensor<int32, [4]> var_101 = const()[name = tensor<string, []>("op_101"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> var_105 = const()[name = tensor<string, []>("op_105"), val = tensor<int32, [4]>([1, 128, 8, 64])];
+            tensor<fp16, [1, 128, 8, 64]> x_11_cast_fp16 = reshape(shape = var_105, x = linear_0_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
+            tensor<bool, []> attention_scores_1_transpose_x_0 = const()[name = tensor<string, []>("attention_scores_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attention_scores_1_transpose_y_0 = const()[name = tensor<string, []>("attention_scores_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_12_perm_0 = const()[name = tensor<string, []>("transpose_12_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_13_perm_0 = const()[name = tensor<string, []>("transpose_13_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 128]> transpose_13 = transpose(perm = transpose_13_perm_0, x = x_3_cast_fp16)[name = tensor<string, []>("transpose_33")];
+            tensor<fp16, [1, 8, 128, 64]> transpose_12 = transpose(perm = transpose_12_perm_0, x = x_11_cast_fp16)[name = tensor<string, []>("transpose_34")];
+            tensor<fp16, [1, 8, 128, 128]> attention_scores_1_cast_fp16 = matmul(transpose_x = attention_scores_1_transpose_x_0, transpose_y = attention_scores_1_transpose_y_0, x = transpose_12, y = transpose_13)[name = tensor<string, []>("attention_scores_1_cast_fp16")];
+            tensor<fp16, []> _inversed_attention_scores_3_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_attention_scores_3_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 128, 128]> _inversed_attention_scores_3_cast_fp16 = mul(x = attention_scores_1_cast_fp16, y = _inversed_attention_scores_3_y_0_to_fp16)[name = tensor<string, []>("_inversed_attention_scores_3_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 128]> attention_scores_5_cast_fp16 = add(x = _inversed_attention_scores_3_cast_fp16, y = attention_mask_cast_fp16)[name = tensor<string, []>("attention_scores_5_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 128]> bias_3_to_fp16 = const()[name = tensor<string, []>("bias_3_to_fp16"), val = tensor<fp16, [1, 8, 128, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32970432)))];
+            tensor<fp16, [1, 8, 128, 128]> input_7_cast_fp16 = add(x = attention_scores_5_cast_fp16, y = bias_3_to_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 128]> input_9_cast_fp16 = softmax(axis = var_16, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
+            tensor<bool, []> context_layer_1_transpose_x_0 = const()[name = tensor<string, []>("context_layer_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> context_layer_1_transpose_y_0 = const()[name = tensor<string, []>("context_layer_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 64]> value_layer_1_cast_fp16 = transpose(perm = var_101, x = x_7_cast_fp16)[name = tensor<string, []>("transpose_35")];
+            tensor<fp16, [1, 8, 128, 64]> context_layer_1_cast_fp16 = matmul(transpose_x = context_layer_1_transpose_x_0, transpose_y = context_layer_1_transpose_y_0, x = input_9_cast_fp16, y = value_layer_1_cast_fp16)[name = tensor<string, []>("context_layer_1_cast_fp16")];
+            tensor<int32, [4]> var_118 = const()[name = tensor<string, []>("op_118"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_123 = const()[name = tensor<string, []>("op_123"), val = tensor<int32, [3]>([1, 128, 512])];
+            tensor<fp16, [1, 128, 8, 64]> var_119_cast_fp16 = transpose(perm = var_118, x = context_layer_1_cast_fp16)[name = tensor<string, []>("transpose_32")];
+            tensor<fp16, [1, 128, 512]> input_11_cast_fp16 = reshape(shape = var_123, x = var_119_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_0_attention_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33232640)))];
+            tensor<fp16, [512]> m_encoder_layer_0_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33756992)))];
+            tensor<fp16, [1, 128, 512]> linear_3_cast_fp16 = linear(bias = m_encoder_layer_0_attention_output_dense_bias_to_fp16, weight = m_encoder_layer_0_attention_output_dense_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
+            tensor<fp16, [1, 128, 512]> input_15_cast_fp16 = add(x = linear_3_cast_fp16, y = input_5_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
+            tensor<int32, [1]> input_17_axes_0 = const()[name = tensor<string, []>("input_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33758080)))];
+            tensor<fp16, [512]> m_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33759168)))];
+            tensor<fp16, [1, 128, 512]> input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = m_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
+            tensor<fp16, [4096, 512]> m_encoder_layer_0_mlp_gated_layers_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_mlp_gated_layers_weight_to_fp16"), val = tensor<fp16, [4096, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33760256)))];
+            tensor<fp16, [4096]> linear_4_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_4_bias_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37954624)))];
+            tensor<fp16, [1, 128, 4096]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = m_encoder_layer_0_mlp_gated_layers_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
+            tensor<int32, [3]> input_19_begin_0 = const()[name = tensor<string, []>("input_19_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> input_19_end_0 = const()[name = tensor<string, []>("input_19_end_0"), val = tensor<int32, [3]>([1, 128, 2048])];
+            tensor<bool, [3]> input_19_end_mask_0 = const()[name = tensor<string, []>("input_19_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 128, 2048]> input_19_cast_fp16 = slice_by_index(begin = input_19_begin_0, end = input_19_end_0, end_mask = input_19_end_mask_0, x = linear_4_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
+            tensor<int32, [3]> non_gated_1_begin_0 = const()[name = tensor<string, []>("non_gated_1_begin_0"), val = tensor<int32, [3]>([0, 0, 2048])];
+            tensor<int32, [3]> non_gated_1_end_0 = const()[name = tensor<string, []>("non_gated_1_end_0"), val = tensor<int32, [3]>([1, 128, 4096])];
+            tensor<bool, [3]> non_gated_1_end_mask_0 = const()[name = tensor<string, []>("non_gated_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 128, 2048]> non_gated_1_cast_fp16 = slice_by_index(begin = non_gated_1_begin_0, end = non_gated_1_end_0, end_mask = non_gated_1_end_mask_0, x = linear_4_cast_fp16)[name = tensor<string, []>("non_gated_1_cast_fp16")];
+            tensor<string, []> var_147_mode_0 = const()[name = tensor<string, []>("op_147_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 128, 2048]> var_147_cast_fp16 = gelu(mode = var_147_mode_0, x = input_19_cast_fp16)[name = tensor<string, []>("op_147_cast_fp16")];
+            tensor<fp16, [1, 128, 2048]> input_21_cast_fp16 = mul(x = var_147_cast_fp16, y = non_gated_1_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
+            tensor<fp16, [512, 2048]> m_encoder_layer_0_mlp_wo_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_mlp_wo_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37962880)))];
+            tensor<fp16, [512]> m_encoder_layer_0_mlp_wo_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_mlp_wo_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40060096)))];
+            tensor<fp16, [1, 128, 512]> linear_5_cast_fp16 = linear(bias = m_encoder_layer_0_mlp_wo_bias_to_fp16, weight = m_encoder_layer_0_mlp_wo_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
+            tensor<fp16, [1, 128, 512]> input_25_cast_fp16 = add(x = linear_5_cast_fp16, y = input_17_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
+            tensor<int32, [1]> input_27_axes_0 = const()[name = tensor<string, []>("input_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_0_mlp_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_mlp_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40061184)))];
+            tensor<fp16, [512]> m_encoder_layer_0_mlp_layernorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_mlp_layernorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40062272)))];
+            tensor<fp16, [1, 128, 512]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = m_encoder_layer_0_mlp_layernorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_0_mlp_layernorm_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_1_attention_self_query_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_query_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40063360)))];
+            tensor<fp16, [512]> m_encoder_layer_1_attention_self_query_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_query_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40587712)))];
+            tensor<fp16, [1, 128, 512]> linear_6_cast_fp16 = linear(bias = m_encoder_layer_1_attention_self_query_bias_to_fp16, weight = m_encoder_layer_1_attention_self_query_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_1_attention_self_key_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_key_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40588800)))];
+            tensor<fp16, [512]> m_encoder_layer_1_attention_self_key_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_key_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41113152)))];
+            tensor<fp16, [1, 128, 512]> linear_7_cast_fp16 = linear(bias = m_encoder_layer_1_attention_self_key_bias_to_fp16, weight = m_encoder_layer_1_attention_self_key_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
+            tensor<int32, [4]> var_173 = const()[name = tensor<string, []>("op_173"), val = tensor<int32, [4]>([1, 128, 8, 64])];
+            tensor<fp16, [1, 128, 8, 64]> x_15_cast_fp16 = reshape(shape = var_173, x = linear_7_cast_fp16)[name = tensor<string, []>("x_15_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_1_attention_self_value_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_value_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41114240)))];
+            tensor<fp16, [512]> m_encoder_layer_1_attention_self_value_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_value_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41638592)))];
+            tensor<fp16, [1, 128, 512]> linear_8_cast_fp16 = linear(bias = m_encoder_layer_1_attention_self_value_bias_to_fp16, weight = m_encoder_layer_1_attention_self_value_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
+            tensor<int32, [4]> var_182 = const()[name = tensor<string, []>("op_182"), val = tensor<int32, [4]>([1, 128, 8, 64])];
+            tensor<fp16, [1, 128, 8, 64]> x_19_cast_fp16 = reshape(shape = var_182, x = linear_8_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")];
+            tensor<int32, [4]> var_184 = const()[name = tensor<string, []>("op_184"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> var_188 = const()[name = tensor<string, []>("op_188"), val = tensor<int32, [4]>([1, 128, 8, 64])];
+            tensor<fp16, [1, 128, 8, 64]> x_23_cast_fp16 = reshape(shape = var_188, x = linear_6_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")];
+            tensor<bool, []> attention_scores_7_transpose_x_0 = const()[name = tensor<string, []>("attention_scores_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attention_scores_7_transpose_y_0 = const()[name = tensor<string, []>("attention_scores_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_14_perm_0 = const()[name = tensor<string, []>("transpose_14_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_15_perm_0 = const()[name = tensor<string, []>("transpose_15_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 128]> transpose_15 = transpose(perm = transpose_15_perm_0, x = x_15_cast_fp16)[name = tensor<string, []>("transpose_29")];
+            tensor<fp16, [1, 8, 128, 64]> transpose_14 = transpose(perm = transpose_14_perm_0, x = x_23_cast_fp16)[name = tensor<string, []>("transpose_30")];
+            tensor<fp16, [1, 8, 128, 128]> attention_scores_7_cast_fp16 = matmul(transpose_x = attention_scores_7_transpose_x_0, transpose_y = attention_scores_7_transpose_y_0, x = transpose_14, y = transpose_15)[name = tensor<string, []>("attention_scores_7_cast_fp16")];
+            tensor<fp16, []> _inversed_attention_scores_9_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_attention_scores_9_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 128, 128]> _inversed_attention_scores_9_cast_fp16 = mul(x = attention_scores_7_cast_fp16, y = _inversed_attention_scores_9_y_0_to_fp16)[name = tensor<string, []>("_inversed_attention_scores_9_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 128]> attention_scores_11_cast_fp16 = add(x = _inversed_attention_scores_9_cast_fp16, y = attention_mask_cast_fp16)[name = tensor<string, []>("attention_scores_11_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 128]> input_29_cast_fp16 = add(x = attention_scores_11_cast_fp16, y = bias_3_to_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 128]> input_31_cast_fp16 = softmax(axis = var_16, x = input_29_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
+            tensor<bool, []> context_layer_5_transpose_x_0 = const()[name = tensor<string, []>("context_layer_5_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> context_layer_5_transpose_y_0 = const()[name = tensor<string, []>("context_layer_5_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 64]> value_layer_3_cast_fp16 = transpose(perm = var_184, x = x_19_cast_fp16)[name = tensor<string, []>("transpose_31")];
+            tensor<fp16, [1, 8, 128, 64]> context_layer_5_cast_fp16 = matmul(transpose_x = context_layer_5_transpose_x_0, transpose_y = context_layer_5_transpose_y_0, x = input_31_cast_fp16, y = value_layer_3_cast_fp16)[name = tensor<string, []>("context_layer_5_cast_fp16")];
+            tensor<int32, [4]> var_201 = const()[name = tensor<string, []>("op_201"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_206 = const()[name = tensor<string, []>("op_206"), val = tensor<int32, [3]>([1, 128, 512])];
+            tensor<fp16, [1, 128, 8, 64]> var_202_cast_fp16 = transpose(perm = var_201, x = context_layer_5_cast_fp16)[name = tensor<string, []>("transpose_28")];
+            tensor<fp16, [1, 128, 512]> input_33_cast_fp16 = reshape(shape = var_206, x = var_202_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_1_attention_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41639680)))];
+            tensor<fp16, [512]> m_encoder_layer_1_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42164032)))];
+            tensor<fp16, [1, 128, 512]> linear_9_cast_fp16 = linear(bias = m_encoder_layer_1_attention_output_dense_bias_to_fp16, weight = m_encoder_layer_1_attention_output_dense_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
+            tensor<fp16, [1, 128, 512]> input_37_cast_fp16 = add(x = linear_9_cast_fp16, y = input_27_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
+            tensor<int32, [1]> input_39_axes_0 = const()[name = tensor<string, []>("input_39_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42165120)))];
+            tensor<fp16, [512]> m_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42166208)))];
+            tensor<fp16, [1, 128, 512]> input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = m_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
+            tensor<fp16, [4096, 512]> m_encoder_layer_1_mlp_gated_layers_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_mlp_gated_layers_weight_to_fp16"), val = tensor<fp16, [4096, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42167296)))];
+            tensor<fp16, [1, 128, 4096]> linear_10_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = m_encoder_layer_1_mlp_gated_layers_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
+            tensor<int32, [3]> input_41_begin_0 = const()[name = tensor<string, []>("input_41_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> input_41_end_0 = const()[name = tensor<string, []>("input_41_end_0"), val = tensor<int32, [3]>([1, 128, 2048])];
+            tensor<bool, [3]> input_41_end_mask_0 = const()[name = tensor<string, []>("input_41_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 128, 2048]> input_41_cast_fp16 = slice_by_index(begin = input_41_begin_0, end = input_41_end_0, end_mask = input_41_end_mask_0, x = linear_10_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
+            tensor<int32, [3]> non_gated_3_begin_0 = const()[name = tensor<string, []>("non_gated_3_begin_0"), val = tensor<int32, [3]>([0, 0, 2048])];
+            tensor<int32, [3]> non_gated_3_end_0 = const()[name = tensor<string, []>("non_gated_3_end_0"), val = tensor<int32, [3]>([1, 128, 4096])];
+            tensor<bool, [3]> non_gated_3_end_mask_0 = const()[name = tensor<string, []>("non_gated_3_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 128, 2048]> non_gated_3_cast_fp16 = slice_by_index(begin = non_gated_3_begin_0, end = non_gated_3_end_0, end_mask = non_gated_3_end_mask_0, x = linear_10_cast_fp16)[name = tensor<string, []>("non_gated_3_cast_fp16")];
+            tensor<string, []> var_230_mode_0 = const()[name = tensor<string, []>("op_230_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 128, 2048]> var_230_cast_fp16 = gelu(mode = var_230_mode_0, x = input_41_cast_fp16)[name = tensor<string, []>("op_230_cast_fp16")];
+            tensor<fp16, [1, 128, 2048]> input_43_cast_fp16 = mul(x = var_230_cast_fp16, y = non_gated_3_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
+            tensor<fp16, [512, 2048]> m_encoder_layer_1_mlp_wo_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_mlp_wo_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(46361664)))];
+            tensor<fp16, [512]> m_encoder_layer_1_mlp_wo_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_mlp_wo_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48458880)))];
+            tensor<fp16, [1, 128, 512]> linear_11_cast_fp16 = linear(bias = m_encoder_layer_1_mlp_wo_bias_to_fp16, weight = m_encoder_layer_1_mlp_wo_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
+            tensor<fp16, [1, 128, 512]> input_47_cast_fp16 = add(x = linear_11_cast_fp16, y = input_39_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
+            tensor<int32, [1]> input_49_axes_0 = const()[name = tensor<string, []>("input_49_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_1_mlp_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_mlp_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48459968)))];
+            tensor<fp16, [512]> m_encoder_layer_1_mlp_layernorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_mlp_layernorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48461056)))];
+            tensor<fp16, [1, 128, 512]> input_49_cast_fp16 = layer_norm(axes = input_49_axes_0, beta = m_encoder_layer_1_mlp_layernorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_1_mlp_layernorm_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_2_attention_self_query_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_query_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48462144)))];
+            tensor<fp16, [512]> m_encoder_layer_2_attention_self_query_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_query_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48986496)))];
+            tensor<fp16, [1, 128, 512]> linear_12_cast_fp16 = linear(bias = m_encoder_layer_2_attention_self_query_bias_to_fp16, weight = m_encoder_layer_2_attention_self_query_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_2_attention_self_key_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_key_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48987584)))];
+            tensor<fp16, [512]> m_encoder_layer_2_attention_self_key_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_key_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49511936)))];
+            tensor<fp16, [1, 128, 512]> linear_13_cast_fp16 = linear(bias = m_encoder_layer_2_attention_self_key_bias_to_fp16, weight = m_encoder_layer_2_attention_self_key_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
+            tensor<int32, [4]> var_256 = const()[name = tensor<string, []>("op_256"), val = tensor<int32, [4]>([1, 128, 8, 64])];
+            tensor<fp16, [1, 128, 8, 64]> x_27_cast_fp16 = reshape(shape = var_256, x = linear_13_cast_fp16)[name = tensor<string, []>("x_27_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_2_attention_self_value_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_value_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49513024)))];
+            tensor<fp16, [512]> m_encoder_layer_2_attention_self_value_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_value_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50037376)))];
+            tensor<fp16, [1, 128, 512]> linear_14_cast_fp16 = linear(bias = m_encoder_layer_2_attention_self_value_bias_to_fp16, weight = m_encoder_layer_2_attention_self_value_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
+            tensor<int32, [4]> var_265 = const()[name = tensor<string, []>("op_265"), val = tensor<int32, [4]>([1, 128, 8, 64])];
+            tensor<fp16, [1, 128, 8, 64]> x_31_cast_fp16 = reshape(shape = var_265, x = linear_14_cast_fp16)[name = tensor<string, []>("x_31_cast_fp16")];
+            tensor<int32, [4]> var_267 = const()[name = tensor<string, []>("op_267"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> var_271 = const()[name = tensor<string, []>("op_271"), val = tensor<int32, [4]>([1, 128, 8, 64])];
+            tensor<fp16, [1, 128, 8, 64]> x_35_cast_fp16 = reshape(shape = var_271, x = linear_12_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")];
+            tensor<bool, []> attention_scores_13_transpose_x_0 = const()[name = tensor<string, []>("attention_scores_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attention_scores_13_transpose_y_0 = const()[name = tensor<string, []>("attention_scores_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_16_perm_0 = const()[name = tensor<string, []>("transpose_16_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_17_perm_0 = const()[name = tensor<string, []>("transpose_17_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 128]> transpose_17 = transpose(perm = transpose_17_perm_0, x = x_27_cast_fp16)[name = tensor<string, []>("transpose_25")];
+            tensor<fp16, [1, 8, 128, 64]> transpose_16 = transpose(perm = transpose_16_perm_0, x = x_35_cast_fp16)[name = tensor<string, []>("transpose_26")];
+            tensor<fp16, [1, 8, 128, 128]> attention_scores_13_cast_fp16 = matmul(transpose_x = attention_scores_13_transpose_x_0, transpose_y = attention_scores_13_transpose_y_0, x = transpose_16, y = transpose_17)[name = tensor<string, []>("attention_scores_13_cast_fp16")];
+            tensor<fp16, []> _inversed_attention_scores_15_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_attention_scores_15_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 128, 128]> _inversed_attention_scores_15_cast_fp16 = mul(x = attention_scores_13_cast_fp16, y = _inversed_attention_scores_15_y_0_to_fp16)[name = tensor<string, []>("_inversed_attention_scores_15_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 128]> attention_scores_17_cast_fp16 = add(x = _inversed_attention_scores_15_cast_fp16, y = attention_mask_cast_fp16)[name = tensor<string, []>("attention_scores_17_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 128]> input_51_cast_fp16 = add(x = attention_scores_17_cast_fp16, y = bias_3_to_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 128]> input_53_cast_fp16 = softmax(axis = var_16, x = input_51_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
+            tensor<bool, []> context_layer_9_transpose_x_0 = const()[name = tensor<string, []>("context_layer_9_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> context_layer_9_transpose_y_0 = const()[name = tensor<string, []>("context_layer_9_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 64]> value_layer_5_cast_fp16 = transpose(perm = var_267, x = x_31_cast_fp16)[name = tensor<string, []>("transpose_27")];
+            tensor<fp16, [1, 8, 128, 64]> context_layer_9_cast_fp16 = matmul(transpose_x = context_layer_9_transpose_x_0, transpose_y = context_layer_9_transpose_y_0, x = input_53_cast_fp16, y = value_layer_5_cast_fp16)[name = tensor<string, []>("context_layer_9_cast_fp16")];
+            tensor<int32, [4]> var_284 = const()[name = tensor<string, []>("op_284"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_289 = const()[name = tensor<string, []>("op_289"), val = tensor<int32, [3]>([1, 128, 512])];
+            tensor<fp16, [1, 128, 8, 64]> var_285_cast_fp16 = transpose(perm = var_284, x = context_layer_9_cast_fp16)[name = tensor<string, []>("transpose_24")];
+            tensor<fp16, [1, 128, 512]> input_55_cast_fp16 = reshape(shape = var_289, x = var_285_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_2_attention_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50038464)))];
+            tensor<fp16, [512]> m_encoder_layer_2_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50562816)))];
+            tensor<fp16, [1, 128, 512]> linear_15_cast_fp16 = linear(bias = m_encoder_layer_2_attention_output_dense_bias_to_fp16, weight = m_encoder_layer_2_attention_output_dense_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("linear_15_cast_fp16")];
+            tensor<fp16, [1, 128, 512]> input_59_cast_fp16 = add(x = linear_15_cast_fp16, y = input_49_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
+            tensor<int32, [1]> input_61_axes_0 = const()[name = tensor<string, []>("input_61_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50563904)))];
+            tensor<fp16, [512]> m_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50564992)))];
+            tensor<fp16, [1, 128, 512]> input_61_cast_fp16 = layer_norm(axes = input_61_axes_0, beta = m_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
+            tensor<fp16, [4096, 512]> m_encoder_layer_2_mlp_gated_layers_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_mlp_gated_layers_weight_to_fp16"), val = tensor<fp16, [4096, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50566080)))];
+            tensor<fp16, [1, 128, 4096]> linear_16_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = m_encoder_layer_2_mlp_gated_layers_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
+            tensor<int32, [3]> input_63_begin_0 = const()[name = tensor<string, []>("input_63_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> input_63_end_0 = const()[name = tensor<string, []>("input_63_end_0"), val = tensor<int32, [3]>([1, 128, 2048])];
+            tensor<bool, [3]> input_63_end_mask_0 = const()[name = tensor<string, []>("input_63_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 128, 2048]> input_63_cast_fp16 = slice_by_index(begin = input_63_begin_0, end = input_63_end_0, end_mask = input_63_end_mask_0, x = linear_16_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
+            tensor<int32, [3]> non_gated_5_begin_0 = const()[name = tensor<string, []>("non_gated_5_begin_0"), val = tensor<int32, [3]>([0, 0, 2048])];
+            tensor<int32, [3]> non_gated_5_end_0 = const()[name = tensor<string, []>("non_gated_5_end_0"), val = tensor<int32, [3]>([1, 128, 4096])];
+            tensor<bool, [3]> non_gated_5_end_mask_0 = const()[name = tensor<string, []>("non_gated_5_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 128, 2048]> non_gated_5_cast_fp16 = slice_by_index(begin = non_gated_5_begin_0, end = non_gated_5_end_0, end_mask = non_gated_5_end_mask_0, x = linear_16_cast_fp16)[name = tensor<string, []>("non_gated_5_cast_fp16")];
+            tensor<string, []> var_313_mode_0 = const()[name = tensor<string, []>("op_313_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 128, 2048]> var_313_cast_fp16 = gelu(mode = var_313_mode_0, x = input_63_cast_fp16)[name = tensor<string, []>("op_313_cast_fp16")];
+            tensor<fp16, [1, 128, 2048]> input_65_cast_fp16 = mul(x = var_313_cast_fp16, y = non_gated_5_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
+            tensor<fp16, [512, 2048]> m_encoder_layer_2_mlp_wo_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_mlp_wo_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54760448)))];
+            tensor<fp16, [512]> m_encoder_layer_2_mlp_wo_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_mlp_wo_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56857664)))];
+            tensor<fp16, [1, 128, 512]> linear_17_cast_fp16 = linear(bias = m_encoder_layer_2_mlp_wo_bias_to_fp16, weight = m_encoder_layer_2_mlp_wo_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
+            tensor<fp16, [1, 128, 512]> input_69_cast_fp16 = add(x = linear_17_cast_fp16, y = input_61_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
+            tensor<int32, [1]> input_71_axes_0 = const()[name = tensor<string, []>("input_71_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_2_mlp_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_mlp_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56858752)))];
+            tensor<fp16, [512]> m_encoder_layer_2_mlp_layernorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_mlp_layernorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56859840)))];
+            tensor<fp16, [1, 128, 512]> input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = m_encoder_layer_2_mlp_layernorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_2_mlp_layernorm_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_3_attention_self_query_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_query_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56860928)))];
+            tensor<fp16, [512]> m_encoder_layer_3_attention_self_query_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_query_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57385280)))];
+            tensor<fp16, [1, 128, 512]> linear_18_cast_fp16 = linear(bias = m_encoder_layer_3_attention_self_query_bias_to_fp16, weight = m_encoder_layer_3_attention_self_query_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_3_attention_self_key_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_key_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57386368)))];
+            tensor<fp16, [512]> m_encoder_layer_3_attention_self_key_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_key_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57910720)))];
+            tensor<fp16, [1, 128, 512]> linear_19_cast_fp16 = linear(bias = m_encoder_layer_3_attention_self_key_bias_to_fp16, weight = m_encoder_layer_3_attention_self_key_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
+            tensor<int32, [4]> var_339 = const()[name = tensor<string, []>("op_339"), val = tensor<int32, [4]>([1, 128, 8, 64])];
+            tensor<fp16, [1, 128, 8, 64]> x_39_cast_fp16 = reshape(shape = var_339, x = linear_19_cast_fp16)[name = tensor<string, []>("x_39_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_3_attention_self_value_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_value_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57911808)))];
+            tensor<fp16, [512]> m_encoder_layer_3_attention_self_value_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_value_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58436160)))];
+            tensor<fp16, [1, 128, 512]> linear_20_cast_fp16 = linear(bias = m_encoder_layer_3_attention_self_value_bias_to_fp16, weight = m_encoder_layer_3_attention_self_value_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
+            tensor<int32, [4]> var_348 = const()[name = tensor<string, []>("op_348"), val = tensor<int32, [4]>([1, 128, 8, 64])];
+            tensor<fp16, [1, 128, 8, 64]> x_43_cast_fp16 = reshape(shape = var_348, x = linear_20_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")];
+            tensor<int32, [4]> var_350 = const()[name = tensor<string, []>("op_350"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> var_354 = const()[name = tensor<string, []>("op_354"), val = tensor<int32, [4]>([1, 128, 8, 64])];
+            tensor<fp16, [1, 128, 8, 64]> x_cast_fp16 = reshape(shape = var_354, x = linear_18_cast_fp16)[name = tensor<string, []>("x_cast_fp16")];
+            tensor<bool, []> attention_scores_19_transpose_x_0 = const()[name = tensor<string, []>("attention_scores_19_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attention_scores_19_transpose_y_0 = const()[name = tensor<string, []>("attention_scores_19_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_18_perm_0 = const()[name = tensor<string, []>("transpose_18_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_19_perm_0 = const()[name = tensor<string, []>("transpose_19_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 128]> transpose_19 = transpose(perm = transpose_19_perm_0, x = x_39_cast_fp16)[name = tensor<string, []>("transpose_21")];
+            tensor<fp16, [1, 8, 128, 64]> transpose_18 = transpose(perm = transpose_18_perm_0, x = x_cast_fp16)[name = tensor<string, []>("transpose_22")];
+            tensor<fp16, [1, 8, 128, 128]> attention_scores_19_cast_fp16 = matmul(transpose_x = attention_scores_19_transpose_x_0, transpose_y = attention_scores_19_transpose_y_0, x = transpose_18, y = transpose_19)[name = tensor<string, []>("attention_scores_19_cast_fp16")];
+            tensor<fp16, []> _inversed_attention_scores_21_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_attention_scores_21_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 128, 128]> _inversed_attention_scores_21_cast_fp16 = mul(x = attention_scores_19_cast_fp16, y = _inversed_attention_scores_21_y_0_to_fp16)[name = tensor<string, []>("_inversed_attention_scores_21_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 128]> attention_scores_cast_fp16 = add(x = _inversed_attention_scores_21_cast_fp16, y = attention_mask_cast_fp16)[name = tensor<string, []>("attention_scores_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 128]> input_73_cast_fp16 = add(x = attention_scores_cast_fp16, y = bias_3_to_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
+            tensor<fp16, [1, 8, 128, 128]> input_75_cast_fp16 = softmax(axis = var_16, x = input_73_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
+            tensor<bool, []> context_layer_13_transpose_x_0 = const()[name = tensor<string, []>("context_layer_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> context_layer_13_transpose_y_0 = const()[name = tensor<string, []>("context_layer_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 128, 64]> value_layer_cast_fp16 = transpose(perm = var_350, x = x_43_cast_fp16)[name = tensor<string, []>("transpose_23")];
+            tensor<fp16, [1, 8, 128, 64]> context_layer_13_cast_fp16 = matmul(transpose_x = context_layer_13_transpose_x_0, transpose_y = context_layer_13_transpose_y_0, x = input_75_cast_fp16, y = value_layer_cast_fp16)[name = tensor<string, []>("context_layer_13_cast_fp16")];
+            tensor<int32, [4]> var_367 = const()[name = tensor<string, []>("op_367"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_372 = const()[name = tensor<string, []>("op_372"), val = tensor<int32, [3]>([1, 128, 512])];
+            tensor<fp16, [1, 128, 8, 64]> var_368_cast_fp16 = transpose(perm = var_367, x = context_layer_13_cast_fp16)[name = tensor<string, []>("transpose_20")];
+            tensor<fp16, [1, 128, 512]> input_77_cast_fp16 = reshape(shape = var_372, x = var_368_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_3_attention_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58437248)))];
+            tensor<fp16, [512]> m_encoder_layer_3_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58961600)))];
+            tensor<fp16, [1, 128, 512]> linear_21_cast_fp16 = linear(bias = m_encoder_layer_3_attention_output_dense_bias_to_fp16, weight = m_encoder_layer_3_attention_output_dense_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
+            tensor<fp16, [1, 128, 512]> input_81_cast_fp16 = add(x = linear_21_cast_fp16, y = input_71_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
+            tensor<int32, [1]> input_83_axes_0 = const()[name = tensor<string, []>("input_83_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58962688)))];
+            tensor<fp16, [512]> m_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58963776)))];
+            tensor<fp16, [1, 128, 512]> input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = m_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
+            tensor<fp16, [4096, 512]> m_encoder_layer_3_mlp_gated_layers_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_mlp_gated_layers_weight_to_fp16"), val = tensor<fp16, [4096, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58964864)))];
+            tensor<fp16, [1, 128, 4096]> linear_22_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = m_encoder_layer_3_mlp_gated_layers_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
+            tensor<int32, [3]> input_85_begin_0 = const()[name = tensor<string, []>("input_85_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> input_85_end_0 = const()[name = tensor<string, []>("input_85_end_0"), val = tensor<int32, [3]>([1, 128, 2048])];
+            tensor<bool, [3]> input_85_end_mask_0 = const()[name = tensor<string, []>("input_85_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 128, 2048]> input_85_cast_fp16 = slice_by_index(begin = input_85_begin_0, end = input_85_end_0, end_mask = input_85_end_mask_0, x = linear_22_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
+            tensor<int32, [3]> non_gated_begin_0 = const()[name = tensor<string, []>("non_gated_begin_0"), val = tensor<int32, [3]>([0, 0, 2048])];
+            tensor<int32, [3]> non_gated_end_0 = const()[name = tensor<string, []>("non_gated_end_0"), val = tensor<int32, [3]>([1, 128, 4096])];
+            tensor<bool, [3]> non_gated_end_mask_0 = const()[name = tensor<string, []>("non_gated_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 128, 2048]> non_gated_cast_fp16 = slice_by_index(begin = non_gated_begin_0, end = non_gated_end_0, end_mask = non_gated_end_mask_0, x = linear_22_cast_fp16)[name = tensor<string, []>("non_gated_cast_fp16")];
+            tensor<string, []> var_396_mode_0 = const()[name = tensor<string, []>("op_396_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 128, 2048]> var_396_cast_fp16 = gelu(mode = var_396_mode_0, x = input_85_cast_fp16)[name = tensor<string, []>("op_396_cast_fp16")];
+            tensor<fp16, [1, 128, 2048]> input_87_cast_fp16 = mul(x = var_396_cast_fp16, y = non_gated_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
+            tensor<fp16, [512, 2048]> m_encoder_layer_3_mlp_wo_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_mlp_wo_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63159232)))];
+            tensor<fp16, [512]> m_encoder_layer_3_mlp_wo_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_mlp_wo_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65256448)))];
+            tensor<fp16, [1, 128, 512]> linear_23_cast_fp16 = linear(bias = m_encoder_layer_3_mlp_wo_bias_to_fp16, weight = m_encoder_layer_3_mlp_wo_weight_to_fp16, x = input_87_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
+            tensor<fp16, [1, 128, 512]> input_91_cast_fp16 = add(x = linear_23_cast_fp16, y = input_83_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
+            tensor<int32, [1]> hidden_states_axes_0 = const()[name = tensor<string, []>("hidden_states_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_3_mlp_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_mlp_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65257536)))];
+            tensor<fp16, [512]> m_encoder_layer_3_mlp_layernorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_mlp_layernorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65258624)))];
+            tensor<fp16, [1, 128, 512]> hidden_states_cast_fp16 = layer_norm(axes = hidden_states_axes_0, beta = m_encoder_layer_3_mlp_layernorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_3_mlp_layernorm_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")];
+            tensor<int32, [1]> var_408_axes_0 = const()[name = tensor<string, []>("op_408_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 128, 1]> var_408 = expand_dims(axes = var_408_axes_0, x = cast_0)[name = tensor<string, []>("op_408")];
+            tensor<string, []> cast_28_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_28_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 128, 1]> var_408_to_fp16 = cast(dtype = cast_28_to_fp16_dtype_0, x = var_408)[name = tensor<string, []>("cast_34")];
+            tensor<fp16, [1, 128, 512]> var_414_cast_fp16 = mul(x = hidden_states_cast_fp16, y = var_408_to_fp16)[name = tensor<string, []>("op_414_cast_fp16")];
+            tensor<int32, [1]> var_419_axes_0 = const()[name = tensor<string, []>("op_419_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, []> var_419_keep_dims_0 = const()[name = tensor<string, []>("op_419_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512]> var_419_cast_fp16 = reduce_sum(axes = var_419_axes_0, keep_dims = var_419_keep_dims_0, x = var_414_cast_fp16)[name = tensor<string, []>("op_419_cast_fp16")];
+            tensor<int32, [1]> var_424_axes_0 = const()[name = tensor<string, []>("op_424_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, []> var_424_keep_dims_0 = const()[name = tensor<string, []>("op_424_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1]> var_424_cast_fp16 = reduce_sum(axes = var_424_axes_0, keep_dims = var_424_keep_dims_0, x = var_408_to_fp16)[name = tensor<string, []>("op_424_cast_fp16")];
+            tensor<fp16, []> var_425_to_fp16 = const()[name = tensor<string, []>("op_425_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, []> const_35_to_fp16 = const()[name = tensor<string, []>("const_35_to_fp16"), val = tensor<fp16, []>(inf)];
+            tensor<fp16, [1, 1]> clip_0_cast_fp16 = clip(alpha = var_425_to_fp16, beta = const_35_to_fp16, x = var_424_cast_fp16)[name = tensor<string, []>("clip_0_cast_fp16")];
+            tensor<fp16, [1, 512]> input_cast_fp16 = real_div(x = var_419_cast_fp16, y = clip_0_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [1]> var_431 = const()[name = tensor<string, []>("op_431"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> var_432 = const()[name = tensor<string, []>("op_432"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 1]> var_434_cast_fp16 = reduce_l2_norm(axes = var_431, keep_dims = var_432, x = input_cast_fp16)[name = tensor<string, []>("op_434_cast_fp16")];
+            tensor<fp16, []> var_435_to_fp16 = const()[name = tensor<string, []>("op_435_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [1, 1]> var_436_cast_fp16 = maximum(x = var_434_cast_fp16, y = var_435_to_fp16)[name = tensor<string, []>("op_436_cast_fp16")];
+            tensor<int32, [2]> denom_reps_0 = const()[name = tensor<string, []>("denom_reps_0"), val = tensor<int32, [2]>([1, 512])];
+            tensor<fp16, [1, 512]> denom_cast_fp16 = tile(reps = denom_reps_0, x = var_436_cast_fp16)[name = tensor<string, []>("denom_cast_fp16")];
+            tensor<fp16, [1, 512]> embedding = real_div(x = input_cast_fp16, y = denom_cast_fp16)[name = tensor<string, []>("op_438_cast_fp16")];
+        } -> (embedding);
+}

jina_embedding_128.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc40c4341c0c984531a2fa3dbedf9b0cd3f82d08425aa1fff22086af03d8672e
+size 65259712

jina_embedding_256.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:657b80f3466d4569028ca537965c0d6df58ecae001dedd420ede7b0424ec2323
+size 243

jina_embedding_256.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77a610ceded9a098314665b89b48380996aec55236a099f37bb239d02639c2f1
+size 318

jina_embedding_256.mlmodelc/metadata.json

@@ -0,0 +1,80 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 512)",
+        "shortDescription" : "",
+        "shape" : "[1, 512]",
+        "name" : "embedding",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.mul" : 10,
+      "Ios17.notEqual" : 1,
+      "Ios17.reshape" : 16,
+      "Ios16.softmax" : 4,
+      "Ios17.matmul" : 8,
+      "Ios17.maximum" : 1,
+      "Ios17.transpose" : 16,
+      "Ios17.expandDims" : 3,
+      "Ios17.add" : 17,
+      "Tile" : 1,
+      "Ios17.sliceByIndex" : 8,
+      "Ios16.reduceSum" : 2,
+      "Ios17.gather" : 1,
+      "Ios17.layerNorm" : 9,
+      "Ios17.clip" : 1,
+      "Ios17.sub" : 1,
+      "Ios16.gelu" : 4,
+      "Ios17.cast" : 4,
+      "Ios17.realDiv" : 2,
+      "Ios17.linear" : 24,
+      "Ios16.reduceL2Norm" : 1
+    },
+    "computePrecision" : "Mixed (Float16, Int16, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.version" : "8.3.0",
+      "com.github.apple.coremltools.source" : "torch==2.5.1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 256)",
+        "shortDescription" : "",
+        "shape" : "[1, 256]",
+        "name" : "input_ids",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "jina_embedding_ane_256",
+    "method" : "predict"
+  }
+]

jina_embedding_256.mlmodelc/model.mil

@@ -0,0 +1,320 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.5.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
+{
+    func main<ios17>(tensor<int32, [1, 256]> input_ids) {
+            tensor<int32, []> var_3 = const()[name = tensor<string, []>("op_3"), val = tensor<int32, []>(0)];
+            tensor<bool, [1, 256]> var_4 = not_equal(x = input_ids, y = var_3)[name = tensor<string, []>("op_4")];
+            tensor<string, []> cast_0_dtype_0 = const()[name = tensor<string, []>("cast_0_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, []> var_16 = const()[name = tensor<string, []>("op_16"), val = tensor<int32, []>(-1)];
+            tensor<int32, [1]> var_41_axes_0 = const()[name = tensor<string, []>("op_41_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 256]> cast_0 = cast(dtype = cast_0_dtype_0, x = var_4)[name = tensor<string, []>("cast_37")];
+            tensor<int32, [1, 1, 256]> var_41 = expand_dims(axes = var_41_axes_0, x = cast_0)[name = tensor<string, []>("op_41")];
+            tensor<int32, [1]> var_42_axes_0 = const()[name = tensor<string, []>("op_42_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<int32, [1, 1, 1, 256]> var_42 = expand_dims(axes = var_42_axes_0, x = var_41)[name = tensor<string, []>("op_42")];
+            tensor<fp16, []> var_21_to_fp16 = const()[name = tensor<string, []>("op_21_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
+            tensor<string, []> cast_1_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 1, 1, 256]> var_42_to_fp16 = cast(dtype = cast_1_to_fp16_dtype_0, x = var_42)[name = tensor<string, []>("cast_36")];
+            tensor<fp16, [1, 1, 1, 256]> var_45_cast_fp16 = sub(x = var_21_to_fp16, y = var_42_to_fp16)[name = tensor<string, []>("op_45_cast_fp16")];
+            tensor<fp16, []> var_46_to_fp16 = const()[name = tensor<string, []>("op_46_to_fp16"), val = tensor<fp16, []>(-0x1.d4cp+14)];
+            tensor<fp16, [1, 1, 1, 256]> attention_mask_cast_fp16 = mul(x = var_45_cast_fp16, y = var_46_to_fp16)[name = tensor<string, []>("attention_mask_cast_fp16")];
+            tensor<int32, []> inputs_embeds_axis_0 = const()[name = tensor<string, []>("inputs_embeds_axis_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> inputs_embeds_batch_dims_0 = const()[name = tensor<string, []>("inputs_embeds_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> inputs_embeds_validate_indices_0 = const()[name = tensor<string, []>("inputs_embeds_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [30528, 512]> m_embeddings_word_embeddings_weight_to_fp16 = const()[name = tensor<string, []>("m_embeddings_word_embeddings_weight_to_fp16"), val = tensor<fp16, [30528, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<string, []> input_ids_to_int16_dtype_0 = const()[name = tensor<string, []>("input_ids_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1, 256]> input_ids_to_int16 = cast(dtype = input_ids_to_int16_dtype_0, x = input_ids)[name = tensor<string, []>("cast_35")];
+            tensor<fp16, [1, 256, 512]> inputs_embeds_cast_fp16_cast_uint16 = gather(axis = inputs_embeds_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = input_ids_to_int16, validate_indices = inputs_embeds_validate_indices_0, x = m_embeddings_word_embeddings_weight_to_fp16)[name = tensor<string, []>("inputs_embeds_cast_fp16_cast_uint16")];
+            tensor<fp16, [1, 256, 512]> token_type_embeddings_1_to_fp16 = const()[name = tensor<string, []>("token_type_embeddings_1_to_fp16"), val = tensor<fp16, [1, 256, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31260800)))];
+            tensor<fp16, [1, 256, 512]> input_3_cast_fp16 = add(x = inputs_embeds_cast_fp16_cast_uint16, y = token_type_embeddings_1_to_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<int32, [1]> input_5_axes_0 = const()[name = tensor<string, []>("input_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_embeddings_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_embeddings_LayerNorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31523008)))];
+            tensor<fp16, [512]> m_embeddings_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_embeddings_LayerNorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31524096)))];
+            tensor<fp16, []> var_18_to_fp16 = const()[name = tensor<string, []>("op_18_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [1, 256, 512]> input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = m_embeddings_LayerNorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_embeddings_LayerNorm_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_0_attention_self_query_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_query_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31525184)))];
+            tensor<fp16, [512]> m_encoder_layer_0_attention_self_query_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_query_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32049536)))];
+            tensor<fp16, [1, 256, 512]> linear_0_cast_fp16 = linear(bias = m_encoder_layer_0_attention_self_query_bias_to_fp16, weight = m_encoder_layer_0_attention_self_query_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_0_attention_self_key_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_key_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32050624)))];
+            tensor<fp16, [512]> m_encoder_layer_0_attention_self_key_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_key_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32574976)))];
+            tensor<fp16, [1, 256, 512]> linear_1_cast_fp16 = linear(bias = m_encoder_layer_0_attention_self_key_bias_to_fp16, weight = m_encoder_layer_0_attention_self_key_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
+            tensor<int32, [4]> var_90 = const()[name = tensor<string, []>("op_90"), val = tensor<int32, [4]>([1, 256, 8, 64])];
+            tensor<fp16, [1, 256, 8, 64]> x_3_cast_fp16 = reshape(shape = var_90, x = linear_1_cast_fp16)[name = tensor<string, []>("x_3_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_0_attention_self_value_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_value_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32576064)))];
+            tensor<fp16, [512]> m_encoder_layer_0_attention_self_value_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_self_value_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33100416)))];
+            tensor<fp16, [1, 256, 512]> linear_2_cast_fp16 = linear(bias = m_encoder_layer_0_attention_self_value_bias_to_fp16, weight = m_encoder_layer_0_attention_self_value_weight_to_fp16, x = input_5_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<int32, [4]> var_99 = const()[name = tensor<string, []>("op_99"), val = tensor<int32, [4]>([1, 256, 8, 64])];
+            tensor<fp16, [1, 256, 8, 64]> x_7_cast_fp16 = reshape(shape = var_99, x = linear_2_cast_fp16)[name = tensor<string, []>("x_7_cast_fp16")];
+            tensor<int32, [4]> var_101 = const()[name = tensor<string, []>("op_101"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> var_105 = const()[name = tensor<string, []>("op_105"), val = tensor<int32, [4]>([1, 256, 8, 64])];
+            tensor<fp16, [1, 256, 8, 64]> x_11_cast_fp16 = reshape(shape = var_105, x = linear_0_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
+            tensor<bool, []> attention_scores_1_transpose_x_0 = const()[name = tensor<string, []>("attention_scores_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attention_scores_1_transpose_y_0 = const()[name = tensor<string, []>("attention_scores_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_12_perm_0 = const()[name = tensor<string, []>("transpose_12_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_13_perm_0 = const()[name = tensor<string, []>("transpose_13_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 256]> transpose_13 = transpose(perm = transpose_13_perm_0, x = x_3_cast_fp16)[name = tensor<string, []>("transpose_33")];
+            tensor<fp16, [1, 8, 256, 64]> transpose_12 = transpose(perm = transpose_12_perm_0, x = x_11_cast_fp16)[name = tensor<string, []>("transpose_34")];
+            tensor<fp16, [1, 8, 256, 256]> attention_scores_1_cast_fp16 = matmul(transpose_x = attention_scores_1_transpose_x_0, transpose_y = attention_scores_1_transpose_y_0, x = transpose_12, y = transpose_13)[name = tensor<string, []>("attention_scores_1_cast_fp16")];
+            tensor<fp16, []> _inversed_attention_scores_3_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_attention_scores_3_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 256, 256]> _inversed_attention_scores_3_cast_fp16 = mul(x = attention_scores_1_cast_fp16, y = _inversed_attention_scores_3_y_0_to_fp16)[name = tensor<string, []>("_inversed_attention_scores_3_cast_fp16")];
+            tensor<fp16, [1, 8, 256, 256]> attention_scores_5_cast_fp16 = add(x = _inversed_attention_scores_3_cast_fp16, y = attention_mask_cast_fp16)[name = tensor<string, []>("attention_scores_5_cast_fp16")];
+            tensor<fp16, [1, 8, 256, 256]> bias_3_to_fp16 = const()[name = tensor<string, []>("bias_3_to_fp16"), val = tensor<fp16, [1, 8, 256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33101504)))];
+            tensor<fp16, [1, 8, 256, 256]> input_7_cast_fp16 = add(x = attention_scores_5_cast_fp16, y = bias_3_to_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<fp16, [1, 8, 256, 256]> input_9_cast_fp16 = softmax(axis = var_16, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
+            tensor<bool, []> context_layer_1_transpose_x_0 = const()[name = tensor<string, []>("context_layer_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> context_layer_1_transpose_y_0 = const()[name = tensor<string, []>("context_layer_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 256, 64]> value_layer_1_cast_fp16 = transpose(perm = var_101, x = x_7_cast_fp16)[name = tensor<string, []>("transpose_35")];
+            tensor<fp16, [1, 8, 256, 64]> context_layer_1_cast_fp16 = matmul(transpose_x = context_layer_1_transpose_x_0, transpose_y = context_layer_1_transpose_y_0, x = input_9_cast_fp16, y = value_layer_1_cast_fp16)[name = tensor<string, []>("context_layer_1_cast_fp16")];
+            tensor<int32, [4]> var_118 = const()[name = tensor<string, []>("op_118"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_123 = const()[name = tensor<string, []>("op_123"), val = tensor<int32, [3]>([1, 256, 512])];
+            tensor<fp16, [1, 256, 8, 64]> var_119_cast_fp16 = transpose(perm = var_118, x = context_layer_1_cast_fp16)[name = tensor<string, []>("transpose_32")];
+            tensor<fp16, [1, 256, 512]> input_11_cast_fp16 = reshape(shape = var_123, x = var_119_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_0_attention_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34150144)))];
+            tensor<fp16, [512]> m_encoder_layer_0_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34674496)))];
+            tensor<fp16, [1, 256, 512]> linear_3_cast_fp16 = linear(bias = m_encoder_layer_0_attention_output_dense_bias_to_fp16, weight = m_encoder_layer_0_attention_output_dense_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
+            tensor<fp16, [1, 256, 512]> input_15_cast_fp16 = add(x = linear_3_cast_fp16, y = input_5_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
+            tensor<int32, [1]> input_17_axes_0 = const()[name = tensor<string, []>("input_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34675584)))];
+            tensor<fp16, [512]> m_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34676672)))];
+            tensor<fp16, [1, 256, 512]> input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = m_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
+            tensor<fp16, [4096, 512]> m_encoder_layer_0_mlp_gated_layers_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_mlp_gated_layers_weight_to_fp16"), val = tensor<fp16, [4096, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34677760)))];
+            tensor<fp16, [4096]> linear_4_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_4_bias_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38872128)))];
+            tensor<fp16, [1, 256, 4096]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = m_encoder_layer_0_mlp_gated_layers_weight_to_fp16, x = input_17_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
+            tensor<int32, [3]> input_19_begin_0 = const()[name = tensor<string, []>("input_19_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> input_19_end_0 = const()[name = tensor<string, []>("input_19_end_0"), val = tensor<int32, [3]>([1, 256, 2048])];
+            tensor<bool, [3]> input_19_end_mask_0 = const()[name = tensor<string, []>("input_19_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 256, 2048]> input_19_cast_fp16 = slice_by_index(begin = input_19_begin_0, end = input_19_end_0, end_mask = input_19_end_mask_0, x = linear_4_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
+            tensor<int32, [3]> non_gated_1_begin_0 = const()[name = tensor<string, []>("non_gated_1_begin_0"), val = tensor<int32, [3]>([0, 0, 2048])];
+            tensor<int32, [3]> non_gated_1_end_0 = const()[name = tensor<string, []>("non_gated_1_end_0"), val = tensor<int32, [3]>([1, 256, 4096])];
+            tensor<bool, [3]> non_gated_1_end_mask_0 = const()[name = tensor<string, []>("non_gated_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 256, 2048]> non_gated_1_cast_fp16 = slice_by_index(begin = non_gated_1_begin_0, end = non_gated_1_end_0, end_mask = non_gated_1_end_mask_0, x = linear_4_cast_fp16)[name = tensor<string, []>("non_gated_1_cast_fp16")];
+            tensor<string, []> var_147_mode_0 = const()[name = tensor<string, []>("op_147_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 256, 2048]> var_147_cast_fp16 = gelu(mode = var_147_mode_0, x = input_19_cast_fp16)[name = tensor<string, []>("op_147_cast_fp16")];
+            tensor<fp16, [1, 256, 2048]> input_21_cast_fp16 = mul(x = var_147_cast_fp16, y = non_gated_1_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
+            tensor<fp16, [512, 2048]> m_encoder_layer_0_mlp_wo_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_mlp_wo_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38880384)))];
+            tensor<fp16, [512]> m_encoder_layer_0_mlp_wo_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_mlp_wo_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40977600)))];
+            tensor<fp16, [1, 256, 512]> linear_5_cast_fp16 = linear(bias = m_encoder_layer_0_mlp_wo_bias_to_fp16, weight = m_encoder_layer_0_mlp_wo_weight_to_fp16, x = input_21_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
+            tensor<fp16, [1, 256, 512]> input_25_cast_fp16 = add(x = linear_5_cast_fp16, y = input_17_cast_fp16)[name = tensor<string, []>("input_25_cast_fp16")];
+            tensor<int32, [1]> input_27_axes_0 = const()[name = tensor<string, []>("input_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_0_mlp_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_mlp_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40978688)))];
+            tensor<fp16, [512]> m_encoder_layer_0_mlp_layernorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_0_mlp_layernorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40979776)))];
+            tensor<fp16, [1, 256, 512]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = m_encoder_layer_0_mlp_layernorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_0_mlp_layernorm_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_1_attention_self_query_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_query_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40980864)))];
+            tensor<fp16, [512]> m_encoder_layer_1_attention_self_query_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_query_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41505216)))];
+            tensor<fp16, [1, 256, 512]> linear_6_cast_fp16 = linear(bias = m_encoder_layer_1_attention_self_query_bias_to_fp16, weight = m_encoder_layer_1_attention_self_query_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_1_attention_self_key_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_key_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41506304)))];
+            tensor<fp16, [512]> m_encoder_layer_1_attention_self_key_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_key_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42030656)))];
+            tensor<fp16, [1, 256, 512]> linear_7_cast_fp16 = linear(bias = m_encoder_layer_1_attention_self_key_bias_to_fp16, weight = m_encoder_layer_1_attention_self_key_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
+            tensor<int32, [4]> var_173 = const()[name = tensor<string, []>("op_173"), val = tensor<int32, [4]>([1, 256, 8, 64])];
+            tensor<fp16, [1, 256, 8, 64]> x_15_cast_fp16 = reshape(shape = var_173, x = linear_7_cast_fp16)[name = tensor<string, []>("x_15_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_1_attention_self_value_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_value_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42031744)))];
+            tensor<fp16, [512]> m_encoder_layer_1_attention_self_value_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_self_value_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42556096)))];
+            tensor<fp16, [1, 256, 512]> linear_8_cast_fp16 = linear(bias = m_encoder_layer_1_attention_self_value_bias_to_fp16, weight = m_encoder_layer_1_attention_self_value_weight_to_fp16, x = input_27_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
+            tensor<int32, [4]> var_182 = const()[name = tensor<string, []>("op_182"), val = tensor<int32, [4]>([1, 256, 8, 64])];
+            tensor<fp16, [1, 256, 8, 64]> x_19_cast_fp16 = reshape(shape = var_182, x = linear_8_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")];
+            tensor<int32, [4]> var_184 = const()[name = tensor<string, []>("op_184"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> var_188 = const()[name = tensor<string, []>("op_188"), val = tensor<int32, [4]>([1, 256, 8, 64])];
+            tensor<fp16, [1, 256, 8, 64]> x_23_cast_fp16 = reshape(shape = var_188, x = linear_6_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")];
+            tensor<bool, []> attention_scores_7_transpose_x_0 = const()[name = tensor<string, []>("attention_scores_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attention_scores_7_transpose_y_0 = const()[name = tensor<string, []>("attention_scores_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_14_perm_0 = const()[name = tensor<string, []>("transpose_14_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_15_perm_0 = const()[name = tensor<string, []>("transpose_15_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 256]> transpose_15 = transpose(perm = transpose_15_perm_0, x = x_15_cast_fp16)[name = tensor<string, []>("transpose_29")];
+            tensor<fp16, [1, 8, 256, 64]> transpose_14 = transpose(perm = transpose_14_perm_0, x = x_23_cast_fp16)[name = tensor<string, []>("transpose_30")];
+            tensor<fp16, [1, 8, 256, 256]> attention_scores_7_cast_fp16 = matmul(transpose_x = attention_scores_7_transpose_x_0, transpose_y = attention_scores_7_transpose_y_0, x = transpose_14, y = transpose_15)[name = tensor<string, []>("attention_scores_7_cast_fp16")];
+            tensor<fp16, []> _inversed_attention_scores_9_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_attention_scores_9_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 256, 256]> _inversed_attention_scores_9_cast_fp16 = mul(x = attention_scores_7_cast_fp16, y = _inversed_attention_scores_9_y_0_to_fp16)[name = tensor<string, []>("_inversed_attention_scores_9_cast_fp16")];
+            tensor<fp16, [1, 8, 256, 256]> attention_scores_11_cast_fp16 = add(x = _inversed_attention_scores_9_cast_fp16, y = attention_mask_cast_fp16)[name = tensor<string, []>("attention_scores_11_cast_fp16")];
+            tensor<fp16, [1, 8, 256, 256]> input_29_cast_fp16 = add(x = attention_scores_11_cast_fp16, y = bias_3_to_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
+            tensor<fp16, [1, 8, 256, 256]> input_31_cast_fp16 = softmax(axis = var_16, x = input_29_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
+            tensor<bool, []> context_layer_5_transpose_x_0 = const()[name = tensor<string, []>("context_layer_5_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> context_layer_5_transpose_y_0 = const()[name = tensor<string, []>("context_layer_5_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 256, 64]> value_layer_3_cast_fp16 = transpose(perm = var_184, x = x_19_cast_fp16)[name = tensor<string, []>("transpose_31")];
+            tensor<fp16, [1, 8, 256, 64]> context_layer_5_cast_fp16 = matmul(transpose_x = context_layer_5_transpose_x_0, transpose_y = context_layer_5_transpose_y_0, x = input_31_cast_fp16, y = value_layer_3_cast_fp16)[name = tensor<string, []>("context_layer_5_cast_fp16")];
+            tensor<int32, [4]> var_201 = const()[name = tensor<string, []>("op_201"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_206 = const()[name = tensor<string, []>("op_206"), val = tensor<int32, [3]>([1, 256, 512])];
+            tensor<fp16, [1, 256, 8, 64]> var_202_cast_fp16 = transpose(perm = var_201, x = context_layer_5_cast_fp16)[name = tensor<string, []>("transpose_28")];
+            tensor<fp16, [1, 256, 512]> input_33_cast_fp16 = reshape(shape = var_206, x = var_202_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_1_attention_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42557184)))];
+            tensor<fp16, [512]> m_encoder_layer_1_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43081536)))];
+            tensor<fp16, [1, 256, 512]> linear_9_cast_fp16 = linear(bias = m_encoder_layer_1_attention_output_dense_bias_to_fp16, weight = m_encoder_layer_1_attention_output_dense_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
+            tensor<fp16, [1, 256, 512]> input_37_cast_fp16 = add(x = linear_9_cast_fp16, y = input_27_cast_fp16)[name = tensor<string, []>("input_37_cast_fp16")];
+            tensor<int32, [1]> input_39_axes_0 = const()[name = tensor<string, []>("input_39_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43082624)))];
+            tensor<fp16, [512]> m_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43083712)))];
+            tensor<fp16, [1, 256, 512]> input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = m_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16, x = input_37_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
+            tensor<fp16, [4096, 512]> m_encoder_layer_1_mlp_gated_layers_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_mlp_gated_layers_weight_to_fp16"), val = tensor<fp16, [4096, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43084800)))];
+            tensor<fp16, [1, 256, 4096]> linear_10_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = m_encoder_layer_1_mlp_gated_layers_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
+            tensor<int32, [3]> input_41_begin_0 = const()[name = tensor<string, []>("input_41_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> input_41_end_0 = const()[name = tensor<string, []>("input_41_end_0"), val = tensor<int32, [3]>([1, 256, 2048])];
+            tensor<bool, [3]> input_41_end_mask_0 = const()[name = tensor<string, []>("input_41_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 256, 2048]> input_41_cast_fp16 = slice_by_index(begin = input_41_begin_0, end = input_41_end_0, end_mask = input_41_end_mask_0, x = linear_10_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")];
+            tensor<int32, [3]> non_gated_3_begin_0 = const()[name = tensor<string, []>("non_gated_3_begin_0"), val = tensor<int32, [3]>([0, 0, 2048])];
+            tensor<int32, [3]> non_gated_3_end_0 = const()[name = tensor<string, []>("non_gated_3_end_0"), val = tensor<int32, [3]>([1, 256, 4096])];
+            tensor<bool, [3]> non_gated_3_end_mask_0 = const()[name = tensor<string, []>("non_gated_3_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 256, 2048]> non_gated_3_cast_fp16 = slice_by_index(begin = non_gated_3_begin_0, end = non_gated_3_end_0, end_mask = non_gated_3_end_mask_0, x = linear_10_cast_fp16)[name = tensor<string, []>("non_gated_3_cast_fp16")];
+            tensor<string, []> var_230_mode_0 = const()[name = tensor<string, []>("op_230_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 256, 2048]> var_230_cast_fp16 = gelu(mode = var_230_mode_0, x = input_41_cast_fp16)[name = tensor<string, []>("op_230_cast_fp16")];
+            tensor<fp16, [1, 256, 2048]> input_43_cast_fp16 = mul(x = var_230_cast_fp16, y = non_gated_3_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
+            tensor<fp16, [512, 2048]> m_encoder_layer_1_mlp_wo_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_mlp_wo_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47279168)))];
+            tensor<fp16, [512]> m_encoder_layer_1_mlp_wo_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_mlp_wo_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49376384)))];
+            tensor<fp16, [1, 256, 512]> linear_11_cast_fp16 = linear(bias = m_encoder_layer_1_mlp_wo_bias_to_fp16, weight = m_encoder_layer_1_mlp_wo_weight_to_fp16, x = input_43_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
+            tensor<fp16, [1, 256, 512]> input_47_cast_fp16 = add(x = linear_11_cast_fp16, y = input_39_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
+            tensor<int32, [1]> input_49_axes_0 = const()[name = tensor<string, []>("input_49_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_1_mlp_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_mlp_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49377472)))];
+            tensor<fp16, [512]> m_encoder_layer_1_mlp_layernorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_1_mlp_layernorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49378560)))];
+            tensor<fp16, [1, 256, 512]> input_49_cast_fp16 = layer_norm(axes = input_49_axes_0, beta = m_encoder_layer_1_mlp_layernorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_1_mlp_layernorm_weight_to_fp16, x = input_47_cast_fp16)[name = tensor<string, []>("input_49_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_2_attention_self_query_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_query_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49379648)))];
+            tensor<fp16, [512]> m_encoder_layer_2_attention_self_query_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_query_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49904000)))];
+            tensor<fp16, [1, 256, 512]> linear_12_cast_fp16 = linear(bias = m_encoder_layer_2_attention_self_query_bias_to_fp16, weight = m_encoder_layer_2_attention_self_query_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_2_attention_self_key_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_key_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49905088)))];
+            tensor<fp16, [512]> m_encoder_layer_2_attention_self_key_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_key_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50429440)))];
+            tensor<fp16, [1, 256, 512]> linear_13_cast_fp16 = linear(bias = m_encoder_layer_2_attention_self_key_bias_to_fp16, weight = m_encoder_layer_2_attention_self_key_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
+            tensor<int32, [4]> var_256 = const()[name = tensor<string, []>("op_256"), val = tensor<int32, [4]>([1, 256, 8, 64])];
+            tensor<fp16, [1, 256, 8, 64]> x_27_cast_fp16 = reshape(shape = var_256, x = linear_13_cast_fp16)[name = tensor<string, []>("x_27_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_2_attention_self_value_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_value_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50430528)))];
+            tensor<fp16, [512]> m_encoder_layer_2_attention_self_value_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_self_value_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50954880)))];
+            tensor<fp16, [1, 256, 512]> linear_14_cast_fp16 = linear(bias = m_encoder_layer_2_attention_self_value_bias_to_fp16, weight = m_encoder_layer_2_attention_self_value_weight_to_fp16, x = input_49_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
+            tensor<int32, [4]> var_265 = const()[name = tensor<string, []>("op_265"), val = tensor<int32, [4]>([1, 256, 8, 64])];
+            tensor<fp16, [1, 256, 8, 64]> x_31_cast_fp16 = reshape(shape = var_265, x = linear_14_cast_fp16)[name = tensor<string, []>("x_31_cast_fp16")];
+            tensor<int32, [4]> var_267 = const()[name = tensor<string, []>("op_267"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> var_271 = const()[name = tensor<string, []>("op_271"), val = tensor<int32, [4]>([1, 256, 8, 64])];
+            tensor<fp16, [1, 256, 8, 64]> x_35_cast_fp16 = reshape(shape = var_271, x = linear_12_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")];
+            tensor<bool, []> attention_scores_13_transpose_x_0 = const()[name = tensor<string, []>("attention_scores_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attention_scores_13_transpose_y_0 = const()[name = tensor<string, []>("attention_scores_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_16_perm_0 = const()[name = tensor<string, []>("transpose_16_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_17_perm_0 = const()[name = tensor<string, []>("transpose_17_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 256]> transpose_17 = transpose(perm = transpose_17_perm_0, x = x_27_cast_fp16)[name = tensor<string, []>("transpose_25")];
+            tensor<fp16, [1, 8, 256, 64]> transpose_16 = transpose(perm = transpose_16_perm_0, x = x_35_cast_fp16)[name = tensor<string, []>("transpose_26")];
+            tensor<fp16, [1, 8, 256, 256]> attention_scores_13_cast_fp16 = matmul(transpose_x = attention_scores_13_transpose_x_0, transpose_y = attention_scores_13_transpose_y_0, x = transpose_16, y = transpose_17)[name = tensor<string, []>("attention_scores_13_cast_fp16")];
+            tensor<fp16, []> _inversed_attention_scores_15_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_attention_scores_15_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 256, 256]> _inversed_attention_scores_15_cast_fp16 = mul(x = attention_scores_13_cast_fp16, y = _inversed_attention_scores_15_y_0_to_fp16)[name = tensor<string, []>("_inversed_attention_scores_15_cast_fp16")];
+            tensor<fp16, [1, 8, 256, 256]> attention_scores_17_cast_fp16 = add(x = _inversed_attention_scores_15_cast_fp16, y = attention_mask_cast_fp16)[name = tensor<string, []>("attention_scores_17_cast_fp16")];
+            tensor<fp16, [1, 8, 256, 256]> input_51_cast_fp16 = add(x = attention_scores_17_cast_fp16, y = bias_3_to_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
+            tensor<fp16, [1, 8, 256, 256]> input_53_cast_fp16 = softmax(axis = var_16, x = input_51_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
+            tensor<bool, []> context_layer_9_transpose_x_0 = const()[name = tensor<string, []>("context_layer_9_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> context_layer_9_transpose_y_0 = const()[name = tensor<string, []>("context_layer_9_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 256, 64]> value_layer_5_cast_fp16 = transpose(perm = var_267, x = x_31_cast_fp16)[name = tensor<string, []>("transpose_27")];
+            tensor<fp16, [1, 8, 256, 64]> context_layer_9_cast_fp16 = matmul(transpose_x = context_layer_9_transpose_x_0, transpose_y = context_layer_9_transpose_y_0, x = input_53_cast_fp16, y = value_layer_5_cast_fp16)[name = tensor<string, []>("context_layer_9_cast_fp16")];
+            tensor<int32, [4]> var_284 = const()[name = tensor<string, []>("op_284"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_289 = const()[name = tensor<string, []>("op_289"), val = tensor<int32, [3]>([1, 256, 512])];
+            tensor<fp16, [1, 256, 8, 64]> var_285_cast_fp16 = transpose(perm = var_284, x = context_layer_9_cast_fp16)[name = tensor<string, []>("transpose_24")];
+            tensor<fp16, [1, 256, 512]> input_55_cast_fp16 = reshape(shape = var_289, x = var_285_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_2_attention_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50955968)))];
+            tensor<fp16, [512]> m_encoder_layer_2_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51480320)))];
+            tensor<fp16, [1, 256, 512]> linear_15_cast_fp16 = linear(bias = m_encoder_layer_2_attention_output_dense_bias_to_fp16, weight = m_encoder_layer_2_attention_output_dense_weight_to_fp16, x = input_55_cast_fp16)[name = tensor<string, []>("linear_15_cast_fp16")];
+            tensor<fp16, [1, 256, 512]> input_59_cast_fp16 = add(x = linear_15_cast_fp16, y = input_49_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
+            tensor<int32, [1]> input_61_axes_0 = const()[name = tensor<string, []>("input_61_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51481408)))];
+            tensor<fp16, [512]> m_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51482496)))];
+            tensor<fp16, [1, 256, 512]> input_61_cast_fp16 = layer_norm(axes = input_61_axes_0, beta = m_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16, x = input_59_cast_fp16)[name = tensor<string, []>("input_61_cast_fp16")];
+            tensor<fp16, [4096, 512]> m_encoder_layer_2_mlp_gated_layers_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_mlp_gated_layers_weight_to_fp16"), val = tensor<fp16, [4096, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(51483584)))];
+            tensor<fp16, [1, 256, 4096]> linear_16_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = m_encoder_layer_2_mlp_gated_layers_weight_to_fp16, x = input_61_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
+            tensor<int32, [3]> input_63_begin_0 = const()[name = tensor<string, []>("input_63_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> input_63_end_0 = const()[name = tensor<string, []>("input_63_end_0"), val = tensor<int32, [3]>([1, 256, 2048])];
+            tensor<bool, [3]> input_63_end_mask_0 = const()[name = tensor<string, []>("input_63_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 256, 2048]> input_63_cast_fp16 = slice_by_index(begin = input_63_begin_0, end = input_63_end_0, end_mask = input_63_end_mask_0, x = linear_16_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
+            tensor<int32, [3]> non_gated_5_begin_0 = const()[name = tensor<string, []>("non_gated_5_begin_0"), val = tensor<int32, [3]>([0, 0, 2048])];
+            tensor<int32, [3]> non_gated_5_end_0 = const()[name = tensor<string, []>("non_gated_5_end_0"), val = tensor<int32, [3]>([1, 256, 4096])];
+            tensor<bool, [3]> non_gated_5_end_mask_0 = const()[name = tensor<string, []>("non_gated_5_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 256, 2048]> non_gated_5_cast_fp16 = slice_by_index(begin = non_gated_5_begin_0, end = non_gated_5_end_0, end_mask = non_gated_5_end_mask_0, x = linear_16_cast_fp16)[name = tensor<string, []>("non_gated_5_cast_fp16")];
+            tensor<string, []> var_313_mode_0 = const()[name = tensor<string, []>("op_313_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 256, 2048]> var_313_cast_fp16 = gelu(mode = var_313_mode_0, x = input_63_cast_fp16)[name = tensor<string, []>("op_313_cast_fp16")];
+            tensor<fp16, [1, 256, 2048]> input_65_cast_fp16 = mul(x = var_313_cast_fp16, y = non_gated_5_cast_fp16)[name = tensor<string, []>("input_65_cast_fp16")];
+            tensor<fp16, [512, 2048]> m_encoder_layer_2_mlp_wo_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_mlp_wo_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(55677952)))];
+            tensor<fp16, [512]> m_encoder_layer_2_mlp_wo_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_mlp_wo_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57775168)))];
+            tensor<fp16, [1, 256, 512]> linear_17_cast_fp16 = linear(bias = m_encoder_layer_2_mlp_wo_bias_to_fp16, weight = m_encoder_layer_2_mlp_wo_weight_to_fp16, x = input_65_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
+            tensor<fp16, [1, 256, 512]> input_69_cast_fp16 = add(x = linear_17_cast_fp16, y = input_61_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
+            tensor<int32, [1]> input_71_axes_0 = const()[name = tensor<string, []>("input_71_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_2_mlp_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_mlp_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57776256)))];
+            tensor<fp16, [512]> m_encoder_layer_2_mlp_layernorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_2_mlp_layernorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57777344)))];
+            tensor<fp16, [1, 256, 512]> input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = m_encoder_layer_2_mlp_layernorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_2_mlp_layernorm_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_3_attention_self_query_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_query_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57778432)))];
+            tensor<fp16, [512]> m_encoder_layer_3_attention_self_query_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_query_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58302784)))];
+            tensor<fp16, [1, 256, 512]> linear_18_cast_fp16 = linear(bias = m_encoder_layer_3_attention_self_query_bias_to_fp16, weight = m_encoder_layer_3_attention_self_query_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_3_attention_self_key_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_key_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58303872)))];
+            tensor<fp16, [512]> m_encoder_layer_3_attention_self_key_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_key_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58828224)))];
+            tensor<fp16, [1, 256, 512]> linear_19_cast_fp16 = linear(bias = m_encoder_layer_3_attention_self_key_bias_to_fp16, weight = m_encoder_layer_3_attention_self_key_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
+            tensor<int32, [4]> var_339 = const()[name = tensor<string, []>("op_339"), val = tensor<int32, [4]>([1, 256, 8, 64])];
+            tensor<fp16, [1, 256, 8, 64]> x_39_cast_fp16 = reshape(shape = var_339, x = linear_19_cast_fp16)[name = tensor<string, []>("x_39_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_3_attention_self_value_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_value_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58829312)))];
+            tensor<fp16, [512]> m_encoder_layer_3_attention_self_value_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_self_value_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59353664)))];
+            tensor<fp16, [1, 256, 512]> linear_20_cast_fp16 = linear(bias = m_encoder_layer_3_attention_self_value_bias_to_fp16, weight = m_encoder_layer_3_attention_self_value_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
+            tensor<int32, [4]> var_348 = const()[name = tensor<string, []>("op_348"), val = tensor<int32, [4]>([1, 256, 8, 64])];
+            tensor<fp16, [1, 256, 8, 64]> x_43_cast_fp16 = reshape(shape = var_348, x = linear_20_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")];
+            tensor<int32, [4]> var_350 = const()[name = tensor<string, []>("op_350"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [4]> var_354 = const()[name = tensor<string, []>("op_354"), val = tensor<int32, [4]>([1, 256, 8, 64])];
+            tensor<fp16, [1, 256, 8, 64]> x_cast_fp16 = reshape(shape = var_354, x = linear_18_cast_fp16)[name = tensor<string, []>("x_cast_fp16")];
+            tensor<bool, []> attention_scores_19_transpose_x_0 = const()[name = tensor<string, []>("attention_scores_19_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> attention_scores_19_transpose_y_0 = const()[name = tensor<string, []>("attention_scores_19_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_18_perm_0 = const()[name = tensor<string, []>("transpose_18_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_19_perm_0 = const()[name = tensor<string, []>("transpose_19_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 8, 64, 256]> transpose_19 = transpose(perm = transpose_19_perm_0, x = x_39_cast_fp16)[name = tensor<string, []>("transpose_21")];
+            tensor<fp16, [1, 8, 256, 64]> transpose_18 = transpose(perm = transpose_18_perm_0, x = x_cast_fp16)[name = tensor<string, []>("transpose_22")];
+            tensor<fp16, [1, 8, 256, 256]> attention_scores_19_cast_fp16 = matmul(transpose_x = attention_scores_19_transpose_x_0, transpose_y = attention_scores_19_transpose_y_0, x = transpose_18, y = transpose_19)[name = tensor<string, []>("attention_scores_19_cast_fp16")];
+            tensor<fp16, []> _inversed_attention_scores_21_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_attention_scores_21_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-3)];
+            tensor<fp16, [1, 8, 256, 256]> _inversed_attention_scores_21_cast_fp16 = mul(x = attention_scores_19_cast_fp16, y = _inversed_attention_scores_21_y_0_to_fp16)[name = tensor<string, []>("_inversed_attention_scores_21_cast_fp16")];
+            tensor<fp16, [1, 8, 256, 256]> attention_scores_cast_fp16 = add(x = _inversed_attention_scores_21_cast_fp16, y = attention_mask_cast_fp16)[name = tensor<string, []>("attention_scores_cast_fp16")];
+            tensor<fp16, [1, 8, 256, 256]> input_73_cast_fp16 = add(x = attention_scores_cast_fp16, y = bias_3_to_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
+            tensor<fp16, [1, 8, 256, 256]> input_75_cast_fp16 = softmax(axis = var_16, x = input_73_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
+            tensor<bool, []> context_layer_13_transpose_x_0 = const()[name = tensor<string, []>("context_layer_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> context_layer_13_transpose_y_0 = const()[name = tensor<string, []>("context_layer_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 8, 256, 64]> value_layer_cast_fp16 = transpose(perm = var_350, x = x_43_cast_fp16)[name = tensor<string, []>("transpose_23")];
+            tensor<fp16, [1, 8, 256, 64]> context_layer_13_cast_fp16 = matmul(transpose_x = context_layer_13_transpose_x_0, transpose_y = context_layer_13_transpose_y_0, x = input_75_cast_fp16, y = value_layer_cast_fp16)[name = tensor<string, []>("context_layer_13_cast_fp16")];
+            tensor<int32, [4]> var_367 = const()[name = tensor<string, []>("op_367"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_372 = const()[name = tensor<string, []>("op_372"), val = tensor<int32, [3]>([1, 256, 512])];
+            tensor<fp16, [1, 256, 8, 64]> var_368_cast_fp16 = transpose(perm = var_367, x = context_layer_13_cast_fp16)[name = tensor<string, []>("transpose_20")];
+            tensor<fp16, [1, 256, 512]> input_77_cast_fp16 = reshape(shape = var_372, x = var_368_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
+            tensor<fp16, [512, 512]> m_encoder_layer_3_attention_output_dense_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59354752)))];
+            tensor<fp16, [512]> m_encoder_layer_3_attention_output_dense_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59879104)))];
+            tensor<fp16, [1, 256, 512]> linear_21_cast_fp16 = linear(bias = m_encoder_layer_3_attention_output_dense_bias_to_fp16, weight = m_encoder_layer_3_attention_output_dense_weight_to_fp16, x = input_77_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
+            tensor<fp16, [1, 256, 512]> input_81_cast_fp16 = add(x = linear_21_cast_fp16, y = input_71_cast_fp16)[name = tensor<string, []>("input_81_cast_fp16")];
+            tensor<int32, [1]> input_83_axes_0 = const()[name = tensor<string, []>("input_83_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59880192)))];
+            tensor<fp16, [512]> m_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59881280)))];
+            tensor<fp16, [1, 256, 512]> input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = m_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16, x = input_81_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
+            tensor<fp16, [4096, 512]> m_encoder_layer_3_mlp_gated_layers_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_mlp_gated_layers_weight_to_fp16"), val = tensor<fp16, [4096, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59882368)))];
+            tensor<fp16, [1, 256, 4096]> linear_22_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = m_encoder_layer_3_mlp_gated_layers_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
+            tensor<int32, [3]> input_85_begin_0 = const()[name = tensor<string, []>("input_85_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> input_85_end_0 = const()[name = tensor<string, []>("input_85_end_0"), val = tensor<int32, [3]>([1, 256, 2048])];
+            tensor<bool, [3]> input_85_end_mask_0 = const()[name = tensor<string, []>("input_85_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
+            tensor<fp16, [1, 256, 2048]> input_85_cast_fp16 = slice_by_index(begin = input_85_begin_0, end = input_85_end_0, end_mask = input_85_end_mask_0, x = linear_22_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")];
+            tensor<int32, [3]> non_gated_begin_0 = const()[name = tensor<string, []>("non_gated_begin_0"), val = tensor<int32, [3]>([0, 0, 2048])];
+            tensor<int32, [3]> non_gated_end_0 = const()[name = tensor<string, []>("non_gated_end_0"), val = tensor<int32, [3]>([1, 256, 4096])];
+            tensor<bool, [3]> non_gated_end_mask_0 = const()[name = tensor<string, []>("non_gated_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp16, [1, 256, 2048]> non_gated_cast_fp16 = slice_by_index(begin = non_gated_begin_0, end = non_gated_end_0, end_mask = non_gated_end_mask_0, x = linear_22_cast_fp16)[name = tensor<string, []>("non_gated_cast_fp16")];
+            tensor<string, []> var_396_mode_0 = const()[name = tensor<string, []>("op_396_mode_0"), val = tensor<string, []>("EXACT")];
+            tensor<fp16, [1, 256, 2048]> var_396_cast_fp16 = gelu(mode = var_396_mode_0, x = input_85_cast_fp16)[name = tensor<string, []>("op_396_cast_fp16")];
+            tensor<fp16, [1, 256, 2048]> input_87_cast_fp16 = mul(x = var_396_cast_fp16, y = non_gated_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
+            tensor<fp16, [512, 2048]> m_encoder_layer_3_mlp_wo_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_mlp_wo_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64076736)))];
+            tensor<fp16, [512]> m_encoder_layer_3_mlp_wo_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_mlp_wo_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66173952)))];
+            tensor<fp16, [1, 256, 512]> linear_23_cast_fp16 = linear(bias = m_encoder_layer_3_mlp_wo_bias_to_fp16, weight = m_encoder_layer_3_mlp_wo_weight_to_fp16, x = input_87_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
+            tensor<fp16, [1, 256, 512]> input_91_cast_fp16 = add(x = linear_23_cast_fp16, y = input_83_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
+            tensor<int32, [1]> hidden_states_axes_0 = const()[name = tensor<string, []>("hidden_states_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [512]> m_encoder_layer_3_mlp_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_mlp_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66175040)))];
+            tensor<fp16, [512]> m_encoder_layer_3_mlp_layernorm_bias_to_fp16 = const()[name = tensor<string, []>("m_encoder_layer_3_mlp_layernorm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66176128)))];
+            tensor<fp16, [1, 256, 512]> hidden_states_cast_fp16 = layer_norm(axes = hidden_states_axes_0, beta = m_encoder_layer_3_mlp_layernorm_bias_to_fp16, epsilon = var_18_to_fp16, gamma = m_encoder_layer_3_mlp_layernorm_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("hidden_states_cast_fp16")];
+            tensor<int32, [1]> var_408_axes_0 = const()[name = tensor<string, []>("op_408_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 256, 1]> var_408 = expand_dims(axes = var_408_axes_0, x = cast_0)[name = tensor<string, []>("op_408")];
+            tensor<string, []> cast_28_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_28_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1, 256, 1]> var_408_to_fp16 = cast(dtype = cast_28_to_fp16_dtype_0, x = var_408)[name = tensor<string, []>("cast_34")];
+            tensor<fp16, [1, 256, 512]> var_414_cast_fp16 = mul(x = hidden_states_cast_fp16, y = var_408_to_fp16)[name = tensor<string, []>("op_414_cast_fp16")];
+            tensor<int32, [1]> var_419_axes_0 = const()[name = tensor<string, []>("op_419_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, []> var_419_keep_dims_0 = const()[name = tensor<string, []>("op_419_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 512]> var_419_cast_fp16 = reduce_sum(axes = var_419_axes_0, keep_dims = var_419_keep_dims_0, x = var_414_cast_fp16)[name = tensor<string, []>("op_419_cast_fp16")];
+            tensor<int32, [1]> var_424_axes_0 = const()[name = tensor<string, []>("op_424_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, []> var_424_keep_dims_0 = const()[name = tensor<string, []>("op_424_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [1, 1]> var_424_cast_fp16 = reduce_sum(axes = var_424_axes_0, keep_dims = var_424_keep_dims_0, x = var_408_to_fp16)[name = tensor<string, []>("op_424_cast_fp16")];
+            tensor<fp16, []> var_425_to_fp16 = const()[name = tensor<string, []>("op_425_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, []> const_35_to_fp16 = const()[name = tensor<string, []>("const_35_to_fp16"), val = tensor<fp16, []>(inf)];
+            tensor<fp16, [1, 1]> clip_0_cast_fp16 = clip(alpha = var_425_to_fp16, beta = const_35_to_fp16, x = var_424_cast_fp16)[name = tensor<string, []>("clip_0_cast_fp16")];
+            tensor<fp16, [1, 512]> input_cast_fp16 = real_div(x = var_419_cast_fp16, y = clip_0_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [1]> var_431 = const()[name = tensor<string, []>("op_431"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> var_432 = const()[name = tensor<string, []>("op_432"), val = tensor<bool, []>(true)];
+            tensor<fp16, [1, 1]> var_434_cast_fp16 = reduce_l2_norm(axes = var_431, keep_dims = var_432, x = input_cast_fp16)[name = tensor<string, []>("op_434_cast_fp16")];
+            tensor<fp16, []> var_435_to_fp16 = const()[name = tensor<string, []>("op_435_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [1, 1]> var_436_cast_fp16 = maximum(x = var_434_cast_fp16, y = var_435_to_fp16)[name = tensor<string, []>("op_436_cast_fp16")];
+            tensor<int32, [2]> denom_reps_0 = const()[name = tensor<string, []>("denom_reps_0"), val = tensor<int32, [2]>([1, 512])];
+            tensor<fp16, [1, 512]> denom_cast_fp16 = tile(reps = denom_reps_0, x = var_436_cast_fp16)[name = tensor<string, []>("denom_cast_fp16")];
+            tensor<fp16, [1, 512]> embedding = real_div(x = input_cast_fp16, y = denom_cast_fp16)[name = tensor<string, []>("op_438_cast_fp16")];
+        } -> (embedding);
+}

jina_embedding_256.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f50af56cb05f6b35c3a37384808e09b1d197d23afe4f6c20d3319754a458a692
+size 66177216