Upload folder using huggingface_hub

manifest.json

@@ -0,0 +1,39 @@
+{
+  "format_version": 1,
+  "source_model": "cross-encoder/ms-marco-MiniLM-L6-v2",
+  "conversion": "Core ML MLProgram FP16 with grouped-channel LUT palettization",
+  "lut_bits": 6,
+  "minimum_macos": "15.0",
+  "preferred_compute_units": "cpuAndNeuralEngine",
+  "input_contract": {
+    "inputs": [
+      "input_ids",
+      "attention_mask",
+      "token_type_ids"
+    ],
+    "output": "score",
+    "dtype": "int32"
+  },
+  "buckets": [
+    {
+      "sequence_length": 128,
+      "artifact": "minilm_reranker_128.mlmodelc",
+      "sha256_tree": "72c11536901e1695ff034d704a58d9bcf9b642e7cd7472daea62b820f3f77ee7",
+      "parity": {
+        "max_absolute_logit_error": 0.19604873657226562,
+        "rank_correlation": 1.0,
+        "identical_probe_order": true
+      }
+    },
+    {
+      "sequence_length": 256,
+      "artifact": "minilm_reranker_256.mlmodelc",
+      "sha256_tree": "17884cc26f17825a80b98e6eb08af08fa85abadb0ae43ba2c19b90a5959358a1",
+      "parity": {
+        "max_absolute_logit_error": 0.20776748657226562,
+        "rank_correlation": 1.0,
+        "identical_probe_order": true
+      }
+    }
+  ]
+}

minilm_reranker_128.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7332a15cde8780f93abf09b371f3f020f9ed3ebdb957e04790e0f198c33fb422
+size 243

minilm_reranker_128.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:456720f72519dfbade860dc610711b677dc7daf0487b3dc531258e3b1fd555c6
+size 434

minilm_reranker_128.mlmodelc/metadata.json

@@ -0,0 +1,97 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Mixed (Float16, Palettized (11 bits), Palettized (13 bits), UInt6)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1]",
+        "name" : "score",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.linear" : 38,
+      "Ios18.constexprLutToDense" : 37,
+      "Ios18.sub" : 1,
+      "Select" : 2,
+      "Ios18.matmul" : 12,
+      "Ios18.gelu" : 6,
+      "Ios18.gather" : 2,
+      "Ios18.add" : 22,
+      "Ios18.tanh" : 1,
+      "Ios18.softmax" : 6,
+      "Ios18.sliceByIndex" : 1,
+      "Ios18.layerNorm" : 13,
+      "Ios18.transpose" : 24,
+      "Ios18.cast" : 7,
+      "Ios18.reshape" : 25,
+      "Ios18.greaterEqual" : 2,
+      "Ios18.mul" : 7
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int16, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "15.0",
+      "tvOS" : "18.0",
+      "visionOS" : "2.0",
+      "watchOS" : "11.0",
+      "iOS" : "18.0",
+      "macCatalyst" : "18.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-07-15",
+      "com.github.apple.coremltools.source" : "torch==2.5.0",
+      "com.github.apple.coremltools.version" : "9.0",
+      "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"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "attention_mask",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 128]",
+        "name" : "token_type_ids",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "minilm_reranker_128",
+    "method" : "predict"
+  }
+]

minilm_reranker_128.mlmodelc/model.mil

@@ -0,0 +1,420 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}})]
+{
+    func main<ios18>(tensor<int32, [1, 128]> attention_mask, tensor<int32, [1, 128]> input_ids, tensor<int32, [1, 128]> token_type_ids) {
+            int32 hidden_1_batch_dims_0 = const()[name = string("hidden_1_batch_dims_0"), val = int32(0)];
+            bool hidden_1_validate_indices_0 = const()[name = string("hidden_1_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [30522, 384]> embeddings_word_embeddings_weight_to_fp16 = const()[name = string("embeddings_word_embeddings_weight_to_fp16"), val = tensor<fp16, [30522, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string input_ids_to_int16_dtype_0 = const()[name = string("input_ids_to_int16_dtype_0"), val = string("int16")];
+            string cast_28_dtype_0 = const()[name = string("cast_28_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
+            tensor<int16, [1, 128]> input_ids_to_int16 = cast(dtype = input_ids_to_int16_dtype_0, x = input_ids)[name = string("cast_7")];
+            tensor<int32, [1, 128]> cast_28 = cast(dtype = cast_28_dtype_0, x = input_ids_to_int16)[name = string("cast_6")];
+            tensor<bool, [1, 128]> greater_equal_0 = greater_equal(x = cast_28, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
+            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(30522)];
+            tensor<int32, [1, 128]> add_0 = add(x = cast_28, y = slice_by_index_0)[name = string("add_0")];
+            tensor<int32, [1, 128]> select_0 = select(a = cast_28, b = add_0, cond = greater_equal_0)[name = string("select_0")];
+            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
+            string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
+            tensor<int16, [1, 128]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_5")];
+            tensor<int32, [1, 128]> cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_4")];
+            tensor<bool, [1, 128]> greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
+            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(30522)];
+            tensor<int32, [1, 128]> add_0_1 = add(x = cast_0, y = slice_by_index_0_1)[name = string("add_0_1")];
+            tensor<int32, [1, 128]> select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")];
+            int32 hidden_1_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("hidden_1_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(0)];
+            tensor<fp16, [1, 128, 384]> hidden_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = hidden_1_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = hidden_1_batch_dims_0, indices = select_0_1, validate_indices = hidden_1_validate_indices_0, x = embeddings_word_embeddings_weight_to_fp16)[name = string("hidden_1_cast_fp16_cast_uint16_cast_uint16")];
+            int32 var_182_axis_0 = const()[name = string("op_182_axis_0"), val = int32(0)];
+            int32 var_182_batch_dims_0 = const()[name = string("op_182_batch_dims_0"), val = int32(0)];
+            bool var_182_validate_indices_0 = const()[name = string("op_182_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [2, 384]> embeddings_token_type_embeddings_weight_to_fp16 = const()[name = string("embeddings_token_type_embeddings_weight_to_fp16"), val = tensor<fp16, [2, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23441024)))];
+            string token_type_ids_to_uint16_dtype_0 = const()[name = string("token_type_ids_to_uint16_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 128]> token_type_ids_to_uint16 = cast(dtype = token_type_ids_to_uint16_dtype_0, x = token_type_ids)[name = string("cast_3")];
+            tensor<fp16, [1, 128, 384]> var_182_cast_fp16_cast_uint16 = gather(axis = var_182_axis_0, batch_dims = var_182_batch_dims_0, indices = token_type_ids_to_uint16, validate_indices = var_182_validate_indices_0, x = embeddings_token_type_embeddings_weight_to_fp16)[name = string("op_182_cast_fp16_cast_uint16")];
+            tensor<fp16, [1, 128, 384]> hidden_cast_fp16 = add(x = hidden_1_cast_fp16_cast_uint16_cast_uint16, y = var_182_cast_fp16_cast_uint16)[name = string("hidden_cast_fp16")];
+            tensor<fp16, [1, 128, 384]> var_188_to_fp16 = const()[name = string("op_188_to_fp16"), val = tensor<fp16, [1, 128, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23442624)))];
+            tensor<fp16, [1, 128, 384]> input_7_cast_fp16 = add(x = hidden_cast_fp16, y = var_188_to_fp16)[name = string("input_7_cast_fp16")];
+            tensor<int32, [1]> input_9_axes_0 = const()[name = string("input_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> embeddings_LayerNorm_weight_to_fp16 = const()[name = string("embeddings_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23540992)))];
+            tensor<fp16, [384]> embeddings_LayerNorm_bias_to_fp16 = const()[name = string("embeddings_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23541824)))];
+            fp16 var_192_to_fp16 = const()[name = string("op_192_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 128, 384]> input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = embeddings_LayerNorm_bias_to_fp16, epsilon = var_192_to_fp16, gamma = embeddings_LayerNorm_weight_to_fp16, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")];
+            fp16 var_206_to_fp16 = const()[name = string("op_206_to_fp16"), val = fp16(0x1p+0)];
+            string var_205_to_fp16_dtype_0 = const()[name = string("op_205_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 128]> attention_mask_to_fp16 = cast(dtype = var_205_to_fp16_dtype_0, x = attention_mask)[name = string("cast_2")];
+            tensor<fp16, [1, 128]> var_208_cast_fp16 = sub(x = var_206_to_fp16, y = attention_mask_to_fp16)[name = string("op_208_cast_fp16")];
+            tensor<int32, [4]> var_213 = const()[name = string("op_213"), val = tensor<int32, [4]>([1, 1, 1, 128])];
+            tensor<fp16, [1, 1, 1, 128]> var_214_cast_fp16 = reshape(shape = var_213, x = var_208_cast_fp16)[name = string("op_214_cast_fp16")];
+            fp16 var_215_to_fp16 = const()[name = string("op_215_to_fp16"), val = fp16(-0x1.388p+13)];
+            tensor<fp16, [1, 1, 1, 128]> additive_mask_cast_fp16 = mul(x = var_214_cast_fp16, y = var_215_to_fp16)[name = string("additive_mask_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_0_attention_self_query_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23542656))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23653312))))[name = string("layers_0_attention_self_query_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_0_attention_self_query_bias_to_fp16 = const()[name = string("layers_0_attention_self_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23656448)))];
+            tensor<fp16, [1, 128, 384]> linear_0_cast_fp16 = linear(bias = layers_0_attention_self_query_bias_to_fp16, weight = layers_0_attention_self_query_weight_to_fp16_palettized, x = input_9_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [4]> var_224 = const()[name = string("op_224"), val = tensor<int32, [4]>([1, 128, 12, 32])];
+            tensor<fp16, [1, 128, 12, 32]> var_225_cast_fp16 = reshape(shape = var_224, x = linear_0_cast_fp16)[name = string("op_225_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_0_attention_self_key_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23657280))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23767936))))[name = string("layers_0_attention_self_key_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_0_attention_self_key_bias_to_fp16 = const()[name = string("layers_0_attention_self_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23771072)))];
+            tensor<fp16, [1, 128, 384]> linear_1_cast_fp16 = linear(bias = layers_0_attention_self_key_bias_to_fp16, weight = layers_0_attention_self_key_weight_to_fp16_palettized, x = input_9_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<int32, [4]> var_239 = const()[name = string("op_239"), val = tensor<int32, [4]>([1, 128, 12, 32])];
+            tensor<fp16, [1, 128, 12, 32]> var_240_cast_fp16 = reshape(shape = var_239, x = linear_1_cast_fp16)[name = string("op_240_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_0_attention_self_value_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23771904))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23882560))))[name = string("layers_0_attention_self_value_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_0_attention_self_value_bias_to_fp16 = const()[name = string("layers_0_attention_self_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23885696)))];
+            tensor<fp16, [1, 128, 384]> linear_2_cast_fp16 = linear(bias = layers_0_attention_self_value_bias_to_fp16, weight = layers_0_attention_self_value_weight_to_fp16_palettized, x = input_9_cast_fp16)[name = string("linear_2_cast_fp16")];
+            tensor<int32, [4]> var_254 = const()[name = string("op_254"), val = tensor<int32, [4]>([1, 128, 12, 32])];
+            tensor<fp16, [1, 128, 12, 32]> var_255_cast_fp16 = reshape(shape = var_254, x = linear_2_cast_fp16)[name = string("op_255_cast_fp16")];
+            tensor<int32, [4]> var_260 = const()[name = string("op_260"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool var_265_transpose_x_0 = const()[name = string("op_265_transpose_x_0"), val = bool(false)];
+            bool var_265_transpose_y_0 = const()[name = string("op_265_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_24_perm_0 = const()[name = string("transpose_24_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_25_perm_0 = const()[name = string("transpose_25_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 12, 32, 128]> transpose_25 = transpose(perm = transpose_25_perm_0, x = var_240_cast_fp16)[name = string("transpose_58")];
+            tensor<fp16, [1, 12, 128, 32]> transpose_24 = transpose(perm = transpose_24_perm_0, x = var_225_cast_fp16)[name = string("transpose_59")];
+            tensor<fp16, [1, 12, 128, 128]> var_265_cast_fp16 = matmul(transpose_x = var_265_transpose_x_0, transpose_y = var_265_transpose_y_0, x = transpose_24, y = transpose_25)[name = string("op_265_cast_fp16")];
+            fp16 var_266_to_fp16 = const()[name = string("op_266_to_fp16"), val = fp16(0x1.6ap-3)];
+            tensor<fp16, [1, 12, 128, 128]> weights_1_cast_fp16 = mul(x = var_265_cast_fp16, y = var_266_to_fp16)[name = string("weights_1_cast_fp16")];
+            tensor<fp16, [1, 12, 128, 128]> var_269_cast_fp16 = add(x = weights_1_cast_fp16, y = additive_mask_cast_fp16)[name = string("op_269_cast_fp16")];
+            int32 var_270 = const()[name = string("op_270"), val = int32(-1)];
+            tensor<fp16, [1, 12, 128, 128]> input_13_cast_fp16 = softmax(axis = var_270, x = var_269_cast_fp16)[name = string("input_13_cast_fp16")];
+            bool var_276_transpose_x_0 = const()[name = string("op_276_transpose_x_0"), val = bool(false)];
+            bool var_276_transpose_y_0 = const()[name = string("op_276_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 128, 32]> value_1_cast_fp16 = transpose(perm = var_260, x = var_255_cast_fp16)[name = string("transpose_57")];
+            tensor<fp16, [1, 12, 128, 32]> var_276_cast_fp16 = matmul(transpose_x = var_276_transpose_x_0, transpose_y = var_276_transpose_y_0, x = input_13_cast_fp16, y = value_1_cast_fp16)[name = string("op_276_cast_fp16")];
+            tensor<int32, [4]> var_281 = const()[name = string("op_281"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_286 = const()[name = string("op_286"), val = tensor<int32, [3]>([1, 128, 384])];
+            tensor<fp16, [1, 128, 12, 32]> var_282_cast_fp16 = transpose(perm = var_281, x = var_276_cast_fp16)[name = string("transpose_56")];
+            tensor<fp16, [1, 128, 384]> input_15_cast_fp16 = reshape(shape = var_286, x = var_282_cast_fp16)[name = string("input_15_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_0_attention_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23886528))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23997184))))[name = string("layers_0_attention_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_0_attention_output_dense_bias_to_fp16 = const()[name = string("layers_0_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24000320)))];
+            tensor<fp16, [1, 128, 384]> linear_3_cast_fp16 = linear(bias = layers_0_attention_output_dense_bias_to_fp16, weight = layers_0_attention_output_dense_weight_to_fp16_palettized, x = input_15_cast_fp16)[name = string("linear_3_cast_fp16")];
+            tensor<fp16, [1, 128, 384]> input_19_cast_fp16 = add(x = linear_3_cast_fp16, y = input_9_cast_fp16)[name = string("input_19_cast_fp16")];
+            tensor<int32, [1]> input_21_axes_0 = const()[name = string("input_21_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_0_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_0_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24001152)))];
+            tensor<fp16, [384]> layers_0_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_0_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24001984)))];
+            fp16 var_289_to_fp16 = const()[name = string("op_289_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 128, 384]> input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = layers_0_attention_output_LayerNorm_bias_to_fp16, epsilon = var_289_to_fp16, gamma = layers_0_attention_output_LayerNorm_weight_to_fp16, x = input_19_cast_fp16)[name = string("input_21_cast_fp16")];
+            tensor<fp16, [1536, 384]> layers_0_intermediate_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24002816))), lut = tensor<fp16, [96, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24445248))))[name = string("layers_0_intermediate_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [1536]> layers_0_intermediate_dense_bias_to_fp16 = const()[name = string("layers_0_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24457600)))];
+            tensor<fp16, [1, 128, 1536]> linear_4_cast_fp16 = linear(bias = layers_0_intermediate_dense_bias_to_fp16, weight = layers_0_intermediate_dense_weight_to_fp16_palettized, x = input_21_cast_fp16)[name = string("linear_4_cast_fp16")];
+            string input_25_mode_0 = const()[name = string("input_25_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 128, 1536]> input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_4_cast_fp16)[name = string("input_25_cast_fp16")];
+            tensor<fp16, [384, 1536]> layers_0_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24460736))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24903168))))[name = string("layers_0_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_0_output_dense_bias_to_fp16 = const()[name = string("layers_0_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24906304)))];
+            tensor<fp16, [1, 128, 384]> linear_5_cast_fp16 = linear(bias = layers_0_output_dense_bias_to_fp16, weight = layers_0_output_dense_weight_to_fp16_palettized, x = input_25_cast_fp16)[name = string("linear_5_cast_fp16")];
+            tensor<fp16, [1, 128, 384]> input_29_cast_fp16 = add(x = linear_5_cast_fp16, y = input_21_cast_fp16)[name = string("input_29_cast_fp16")];
+            tensor<int32, [1]> input_31_axes_0 = const()[name = string("input_31_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_0_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_0_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24907136)))];
+            tensor<fp16, [384]> layers_0_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_0_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24907968)))];
+            fp16 var_312_to_fp16 = const()[name = string("op_312_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 128, 384]> input_31_cast_fp16 = layer_norm(axes = input_31_axes_0, beta = layers_0_output_LayerNorm_bias_to_fp16, epsilon = var_312_to_fp16, gamma = layers_0_output_LayerNorm_weight_to_fp16, x = input_29_cast_fp16)[name = string("input_31_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_1_attention_self_query_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24908800))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25019456))))[name = string("layers_1_attention_self_query_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_1_attention_self_query_bias_to_fp16 = const()[name = string("layers_1_attention_self_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25022592)))];
+            tensor<fp16, [1, 128, 384]> linear_6_cast_fp16 = linear(bias = layers_1_attention_self_query_bias_to_fp16, weight = layers_1_attention_self_query_weight_to_fp16_palettized, x = input_31_cast_fp16)[name = string("linear_6_cast_fp16")];
+            tensor<int32, [4]> var_335 = const()[name = string("op_335"), val = tensor<int32, [4]>([1, 128, 12, 32])];
+            tensor<fp16, [1, 128, 12, 32]> var_336_cast_fp16 = reshape(shape = var_335, x = linear_6_cast_fp16)[name = string("op_336_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_1_attention_self_key_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25023424))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25134080))))[name = string("layers_1_attention_self_key_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_1_attention_self_key_bias_to_fp16 = const()[name = string("layers_1_attention_self_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25137216)))];
+            tensor<fp16, [1, 128, 384]> linear_7_cast_fp16 = linear(bias = layers_1_attention_self_key_bias_to_fp16, weight = layers_1_attention_self_key_weight_to_fp16_palettized, x = input_31_cast_fp16)[name = string("linear_7_cast_fp16")];
+            tensor<int32, [4]> var_350 = const()[name = string("op_350"), val = tensor<int32, [4]>([1, 128, 12, 32])];
+            tensor<fp16, [1, 128, 12, 32]> var_351_cast_fp16 = reshape(shape = var_350, x = linear_7_cast_fp16)[name = string("op_351_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_1_attention_self_value_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25138048))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25248704))))[name = string("layers_1_attention_self_value_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_1_attention_self_value_bias_to_fp16 = const()[name = string("layers_1_attention_self_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25251840)))];
+            tensor<fp16, [1, 128, 384]> linear_8_cast_fp16 = linear(bias = layers_1_attention_self_value_bias_to_fp16, weight = layers_1_attention_self_value_weight_to_fp16_palettized, x = input_31_cast_fp16)[name = string("linear_8_cast_fp16")];
+            tensor<int32, [4]> var_365 = const()[name = string("op_365"), val = tensor<int32, [4]>([1, 128, 12, 32])];
+            tensor<fp16, [1, 128, 12, 32]> var_366_cast_fp16 = reshape(shape = var_365, x = linear_8_cast_fp16)[name = string("op_366_cast_fp16")];
+            tensor<int32, [4]> var_371 = const()[name = string("op_371"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool var_376_transpose_x_0 = const()[name = string("op_376_transpose_x_0"), val = bool(false)];
+            bool var_376_transpose_y_0 = const()[name = string("op_376_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_26_perm_0 = const()[name = string("transpose_26_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_27_perm_0 = const()[name = string("transpose_27_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 12, 32, 128]> transpose_27 = transpose(perm = transpose_27_perm_0, x = var_351_cast_fp16)[name = string("transpose_54")];
+            tensor<fp16, [1, 12, 128, 32]> transpose_26 = transpose(perm = transpose_26_perm_0, x = var_336_cast_fp16)[name = string("transpose_55")];
+            tensor<fp16, [1, 12, 128, 128]> var_376_cast_fp16 = matmul(transpose_x = var_376_transpose_x_0, transpose_y = var_376_transpose_y_0, x = transpose_26, y = transpose_27)[name = string("op_376_cast_fp16")];
+            fp16 var_377_to_fp16 = const()[name = string("op_377_to_fp16"), val = fp16(0x1.6ap-3)];
+            tensor<fp16, [1, 12, 128, 128]> weights_5_cast_fp16 = mul(x = var_376_cast_fp16, y = var_377_to_fp16)[name = string("weights_5_cast_fp16")];
+            tensor<fp16, [1, 12, 128, 128]> var_380_cast_fp16 = add(x = weights_5_cast_fp16, y = additive_mask_cast_fp16)[name = string("op_380_cast_fp16")];
+            int32 var_381 = const()[name = string("op_381"), val = int32(-1)];
+            tensor<fp16, [1, 12, 128, 128]> input_33_cast_fp16 = softmax(axis = var_381, x = var_380_cast_fp16)[name = string("input_33_cast_fp16")];
+            bool var_387_transpose_x_0 = const()[name = string("op_387_transpose_x_0"), val = bool(false)];
+            bool var_387_transpose_y_0 = const()[name = string("op_387_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 128, 32]> value_3_cast_fp16 = transpose(perm = var_371, x = var_366_cast_fp16)[name = string("transpose_53")];
+            tensor<fp16, [1, 12, 128, 32]> var_387_cast_fp16 = matmul(transpose_x = var_387_transpose_x_0, transpose_y = var_387_transpose_y_0, x = input_33_cast_fp16, y = value_3_cast_fp16)[name = string("op_387_cast_fp16")];
+            tensor<int32, [4]> var_392 = const()[name = string("op_392"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_397 = const()[name = string("op_397"), val = tensor<int32, [3]>([1, 128, 384])];
+            tensor<fp16, [1, 128, 12, 32]> var_393_cast_fp16 = transpose(perm = var_392, x = var_387_cast_fp16)[name = string("transpose_52")];
+            tensor<fp16, [1, 128, 384]> input_35_cast_fp16 = reshape(shape = var_397, x = var_393_cast_fp16)[name = string("input_35_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_1_attention_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25252672))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25363328))))[name = string("layers_1_attention_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_1_attention_output_dense_bias_to_fp16 = const()[name = string("layers_1_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25366464)))];
+            tensor<fp16, [1, 128, 384]> linear_9_cast_fp16 = linear(bias = layers_1_attention_output_dense_bias_to_fp16, weight = layers_1_attention_output_dense_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("linear_9_cast_fp16")];
+            tensor<fp16, [1, 128, 384]> input_39_cast_fp16 = add(x = linear_9_cast_fp16, y = input_31_cast_fp16)[name = string("input_39_cast_fp16")];
+            tensor<int32, [1]> input_41_axes_0 = const()[name = string("input_41_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_1_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_1_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25367296)))];
+            tensor<fp16, [384]> layers_1_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_1_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25368128)))];
+            fp16 var_400_to_fp16 = const()[name = string("op_400_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 128, 384]> input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = layers_1_attention_output_LayerNorm_bias_to_fp16, epsilon = var_400_to_fp16, gamma = layers_1_attention_output_LayerNorm_weight_to_fp16, x = input_39_cast_fp16)[name = string("input_41_cast_fp16")];
+            tensor<fp16, [1536, 384]> layers_1_intermediate_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25368960))), lut = tensor<fp16, [96, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25811392))))[name = string("layers_1_intermediate_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [1536]> layers_1_intermediate_dense_bias_to_fp16 = const()[name = string("layers_1_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25823744)))];
+            tensor<fp16, [1, 128, 1536]> linear_10_cast_fp16 = linear(bias = layers_1_intermediate_dense_bias_to_fp16, weight = layers_1_intermediate_dense_weight_to_fp16_palettized, x = input_41_cast_fp16)[name = string("linear_10_cast_fp16")];
+            string input_45_mode_0 = const()[name = string("input_45_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 128, 1536]> input_45_cast_fp16 = gelu(mode = input_45_mode_0, x = linear_10_cast_fp16)[name = string("input_45_cast_fp16")];
+            tensor<fp16, [384, 1536]> layers_1_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25826880))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26269312))))[name = string("layers_1_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_1_output_dense_bias_to_fp16 = const()[name = string("layers_1_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26272448)))];
+            tensor<fp16, [1, 128, 384]> linear_11_cast_fp16 = linear(bias = layers_1_output_dense_bias_to_fp16, weight = layers_1_output_dense_weight_to_fp16_palettized, x = input_45_cast_fp16)[name = string("linear_11_cast_fp16")];
+            tensor<fp16, [1, 128, 384]> input_49_cast_fp16 = add(x = linear_11_cast_fp16, y = input_41_cast_fp16)[name = string("input_49_cast_fp16")];
+            tensor<int32, [1]> input_51_axes_0 = const()[name = string("input_51_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_1_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_1_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26273280)))];
+            tensor<fp16, [384]> layers_1_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_1_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26274112)))];
+            fp16 var_423_to_fp16 = const()[name = string("op_423_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 128, 384]> input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = layers_1_output_LayerNorm_bias_to_fp16, epsilon = var_423_to_fp16, gamma = layers_1_output_LayerNorm_weight_to_fp16, x = input_49_cast_fp16)[name = string("input_51_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_2_attention_self_query_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26274944))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26385600))))[name = string("layers_2_attention_self_query_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_2_attention_self_query_bias_to_fp16 = const()[name = string("layers_2_attention_self_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26388736)))];
+            tensor<fp16, [1, 128, 384]> linear_12_cast_fp16 = linear(bias = layers_2_attention_self_query_bias_to_fp16, weight = layers_2_attention_self_query_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = string("linear_12_cast_fp16")];
+            tensor<int32, [4]> var_446 = const()[name = string("op_446"), val = tensor<int32, [4]>([1, 128, 12, 32])];
+            tensor<fp16, [1, 128, 12, 32]> var_447_cast_fp16 = reshape(shape = var_446, x = linear_12_cast_fp16)[name = string("op_447_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_2_attention_self_key_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26389568))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26500224))))[name = string("layers_2_attention_self_key_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_2_attention_self_key_bias_to_fp16 = const()[name = string("layers_2_attention_self_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26503360)))];
+            tensor<fp16, [1, 128, 384]> linear_13_cast_fp16 = linear(bias = layers_2_attention_self_key_bias_to_fp16, weight = layers_2_attention_self_key_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = string("linear_13_cast_fp16")];
+            tensor<int32, [4]> var_461 = const()[name = string("op_461"), val = tensor<int32, [4]>([1, 128, 12, 32])];
+            tensor<fp16, [1, 128, 12, 32]> var_462_cast_fp16 = reshape(shape = var_461, x = linear_13_cast_fp16)[name = string("op_462_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_2_attention_self_value_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26504192))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26614848))))[name = string("layers_2_attention_self_value_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_2_attention_self_value_bias_to_fp16 = const()[name = string("layers_2_attention_self_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26617984)))];
+            tensor<fp16, [1, 128, 384]> linear_14_cast_fp16 = linear(bias = layers_2_attention_self_value_bias_to_fp16, weight = layers_2_attention_self_value_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = string("linear_14_cast_fp16")];
+            tensor<int32, [4]> var_476 = const()[name = string("op_476"), val = tensor<int32, [4]>([1, 128, 12, 32])];
+            tensor<fp16, [1, 128, 12, 32]> var_477_cast_fp16 = reshape(shape = var_476, x = linear_14_cast_fp16)[name = string("op_477_cast_fp16")];
+            tensor<int32, [4]> var_482 = const()[name = string("op_482"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool var_487_transpose_x_0 = const()[name = string("op_487_transpose_x_0"), val = bool(false)];
+            bool var_487_transpose_y_0 = const()[name = string("op_487_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_28_perm_0 = const()[name = string("transpose_28_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_29_perm_0 = const()[name = string("transpose_29_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 12, 32, 128]> transpose_29 = transpose(perm = transpose_29_perm_0, x = var_462_cast_fp16)[name = string("transpose_50")];
+            tensor<fp16, [1, 12, 128, 32]> transpose_28 = transpose(perm = transpose_28_perm_0, x = var_447_cast_fp16)[name = string("transpose_51")];
+            tensor<fp16, [1, 12, 128, 128]> var_487_cast_fp16 = matmul(transpose_x = var_487_transpose_x_0, transpose_y = var_487_transpose_y_0, x = transpose_28, y = transpose_29)[name = string("op_487_cast_fp16")];
+            fp16 var_488_to_fp16 = const()[name = string("op_488_to_fp16"), val = fp16(0x1.6ap-3)];
+            tensor<fp16, [1, 12, 128, 128]> weights_9_cast_fp16 = mul(x = var_487_cast_fp16, y = var_488_to_fp16)[name = string("weights_9_cast_fp16")];
+            tensor<fp16, [1, 12, 128, 128]> var_491_cast_fp16 = add(x = weights_9_cast_fp16, y = additive_mask_cast_fp16)[name = string("op_491_cast_fp16")];
+            int32 var_492 = const()[name = string("op_492"), val = int32(-1)];
+            tensor<fp16, [1, 12, 128, 128]> input_53_cast_fp16 = softmax(axis = var_492, x = var_491_cast_fp16)[name = string("input_53_cast_fp16")];
+            bool var_498_transpose_x_0 = const()[name = string("op_498_transpose_x_0"), val = bool(false)];
+            bool var_498_transpose_y_0 = const()[name = string("op_498_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 128, 32]> value_5_cast_fp16 = transpose(perm = var_482, x = var_477_cast_fp16)[name = string("transpose_49")];
+            tensor<fp16, [1, 12, 128, 32]> var_498_cast_fp16 = matmul(transpose_x = var_498_transpose_x_0, transpose_y = var_498_transpose_y_0, x = input_53_cast_fp16, y = value_5_cast_fp16)[name = string("op_498_cast_fp16")];
+            tensor<int32, [4]> var_503 = const()[name = string("op_503"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_508 = const()[name = string("op_508"), val = tensor<int32, [3]>([1, 128, 384])];
+            tensor<fp16, [1, 128, 12, 32]> var_504_cast_fp16 = transpose(perm = var_503, x = var_498_cast_fp16)[name = string("transpose_48")];
+            tensor<fp16, [1, 128, 384]> input_55_cast_fp16 = reshape(shape = var_508, x = var_504_cast_fp16)[name = string("input_55_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_2_attention_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26618816))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26729472))))[name = string("layers_2_attention_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_2_attention_output_dense_bias_to_fp16 = const()[name = string("layers_2_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26732608)))];
+            tensor<fp16, [1, 128, 384]> linear_15_cast_fp16 = linear(bias = layers_2_attention_output_dense_bias_to_fp16, weight = layers_2_attention_output_dense_weight_to_fp16_palettized, x = input_55_cast_fp16)[name = string("linear_15_cast_fp16")];
+            tensor<fp16, [1, 128, 384]> input_59_cast_fp16 = add(x = linear_15_cast_fp16, y = input_51_cast_fp16)[name = string("input_59_cast_fp16")];
+            tensor<int32, [1]> input_61_axes_0 = const()[name = string("input_61_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_2_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_2_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26733440)))];
+            tensor<fp16, [384]> layers_2_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_2_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26734272)))];
+            fp16 var_511_to_fp16 = const()[name = string("op_511_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 128, 384]> input_61_cast_fp16 = layer_norm(axes = input_61_axes_0, beta = layers_2_attention_output_LayerNorm_bias_to_fp16, epsilon = var_511_to_fp16, gamma = layers_2_attention_output_LayerNorm_weight_to_fp16, x = input_59_cast_fp16)[name = string("input_61_cast_fp16")];
+            tensor<fp16, [1536, 384]> layers_2_intermediate_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26735104))), lut = tensor<fp16, [96, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27177536))))[name = string("layers_2_intermediate_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [1536]> layers_2_intermediate_dense_bias_to_fp16 = const()[name = string("layers_2_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27189888)))];
+            tensor<fp16, [1, 128, 1536]> linear_16_cast_fp16 = linear(bias = layers_2_intermediate_dense_bias_to_fp16, weight = layers_2_intermediate_dense_weight_to_fp16_palettized, x = input_61_cast_fp16)[name = string("linear_16_cast_fp16")];
+            string input_65_mode_0 = const()[name = string("input_65_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 128, 1536]> input_65_cast_fp16 = gelu(mode = input_65_mode_0, x = linear_16_cast_fp16)[name = string("input_65_cast_fp16")];
+            tensor<fp16, [384, 1536]> layers_2_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27193024))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27635456))))[name = string("layers_2_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_2_output_dense_bias_to_fp16 = const()[name = string("layers_2_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27638592)))];
+            tensor<fp16, [1, 128, 384]> linear_17_cast_fp16 = linear(bias = layers_2_output_dense_bias_to_fp16, weight = layers_2_output_dense_weight_to_fp16_palettized, x = input_65_cast_fp16)[name = string("linear_17_cast_fp16")];
+            tensor<fp16, [1, 128, 384]> input_69_cast_fp16 = add(x = linear_17_cast_fp16, y = input_61_cast_fp16)[name = string("input_69_cast_fp16")];
+            tensor<int32, [1]> input_71_axes_0 = const()[name = string("input_71_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_2_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_2_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27639424)))];
+            tensor<fp16, [384]> layers_2_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_2_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27640256)))];
+            fp16 var_534_to_fp16 = const()[name = string("op_534_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 128, 384]> input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = layers_2_output_LayerNorm_bias_to_fp16, epsilon = var_534_to_fp16, gamma = layers_2_output_LayerNorm_weight_to_fp16, x = input_69_cast_fp16)[name = string("input_71_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_3_attention_self_query_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27641088))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27751744))))[name = string("layers_3_attention_self_query_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_3_attention_self_query_bias_to_fp16 = const()[name = string("layers_3_attention_self_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27754880)))];
+            tensor<fp16, [1, 128, 384]> linear_18_cast_fp16 = linear(bias = layers_3_attention_self_query_bias_to_fp16, weight = layers_3_attention_self_query_weight_to_fp16_palettized, x = input_71_cast_fp16)[name = string("linear_18_cast_fp16")];
+            tensor<int32, [4]> var_557 = const()[name = string("op_557"), val = tensor<int32, [4]>([1, 128, 12, 32])];
+            tensor<fp16, [1, 128, 12, 32]> var_558_cast_fp16 = reshape(shape = var_557, x = linear_18_cast_fp16)[name = string("op_558_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_3_attention_self_key_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27755712))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27866368))))[name = string("layers_3_attention_self_key_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_3_attention_self_key_bias_to_fp16 = const()[name = string("layers_3_attention_self_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27869504)))];
+            tensor<fp16, [1, 128, 384]> linear_19_cast_fp16 = linear(bias = layers_3_attention_self_key_bias_to_fp16, weight = layers_3_attention_self_key_weight_to_fp16_palettized, x = input_71_cast_fp16)[name = string("linear_19_cast_fp16")];
+            tensor<int32, [4]> var_572 = const()[name = string("op_572"), val = tensor<int32, [4]>([1, 128, 12, 32])];
+            tensor<fp16, [1, 128, 12, 32]> var_573_cast_fp16 = reshape(shape = var_572, x = linear_19_cast_fp16)[name = string("op_573_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_3_attention_self_value_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27870336))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27980992))))[name = string("layers_3_attention_self_value_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_3_attention_self_value_bias_to_fp16 = const()[name = string("layers_3_attention_self_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27984128)))];
+            tensor<fp16, [1, 128, 384]> linear_20_cast_fp16 = linear(bias = layers_3_attention_self_value_bias_to_fp16, weight = layers_3_attention_self_value_weight_to_fp16_palettized, x = input_71_cast_fp16)[name = string("linear_20_cast_fp16")];
+            tensor<int32, [4]> var_587 = const()[name = string("op_587"), val = tensor<int32, [4]>([1, 128, 12, 32])];
+            tensor<fp16, [1, 128, 12, 32]> var_588_cast_fp16 = reshape(shape = var_587, x = linear_20_cast_fp16)[name = string("op_588_cast_fp16")];
+            tensor<int32, [4]> var_593 = const()[name = string("op_593"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool var_598_transpose_x_0 = const()[name = string("op_598_transpose_x_0"), val = bool(false)];
+            bool var_598_transpose_y_0 = const()[name = string("op_598_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_30_perm_0 = const()[name = string("transpose_30_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_31_perm_0 = const()[name = string("transpose_31_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 12, 32, 128]> transpose_31 = transpose(perm = transpose_31_perm_0, x = var_573_cast_fp16)[name = string("transpose_46")];
+            tensor<fp16, [1, 12, 128, 32]> transpose_30 = transpose(perm = transpose_30_perm_0, x = var_558_cast_fp16)[name = string("transpose_47")];
+            tensor<fp16, [1, 12, 128, 128]> var_598_cast_fp16 = matmul(transpose_x = var_598_transpose_x_0, transpose_y = var_598_transpose_y_0, x = transpose_30, y = transpose_31)[name = string("op_598_cast_fp16")];
+            fp16 var_599_to_fp16 = const()[name = string("op_599_to_fp16"), val = fp16(0x1.6ap-3)];
+            tensor<fp16, [1, 12, 128, 128]> weights_13_cast_fp16 = mul(x = var_598_cast_fp16, y = var_599_to_fp16)[name = string("weights_13_cast_fp16")];
+            tensor<fp16, [1, 12, 128, 128]> var_602_cast_fp16 = add(x = weights_13_cast_fp16, y = additive_mask_cast_fp16)[name = string("op_602_cast_fp16")];
+            int32 var_603 = const()[name = string("op_603"), val = int32(-1)];
+            tensor<fp16, [1, 12, 128, 128]> input_73_cast_fp16 = softmax(axis = var_603, x = var_602_cast_fp16)[name = string("input_73_cast_fp16")];
+            bool var_609_transpose_x_0 = const()[name = string("op_609_transpose_x_0"), val = bool(false)];
+            bool var_609_transpose_y_0 = const()[name = string("op_609_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 128, 32]> value_7_cast_fp16 = transpose(perm = var_593, x = var_588_cast_fp16)[name = string("transpose_45")];
+            tensor<fp16, [1, 12, 128, 32]> var_609_cast_fp16 = matmul(transpose_x = var_609_transpose_x_0, transpose_y = var_609_transpose_y_0, x = input_73_cast_fp16, y = value_7_cast_fp16)[name = string("op_609_cast_fp16")];
+            tensor<int32, [4]> var_614 = const()[name = string("op_614"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_619 = const()[name = string("op_619"), val = tensor<int32, [3]>([1, 128, 384])];
+            tensor<fp16, [1, 128, 12, 32]> var_615_cast_fp16 = transpose(perm = var_614, x = var_609_cast_fp16)[name = string("transpose_44")];
+            tensor<fp16, [1, 128, 384]> input_75_cast_fp16 = reshape(shape = var_619, x = var_615_cast_fp16)[name = string("input_75_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_3_attention_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27984960))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28095616))))[name = string("layers_3_attention_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_3_attention_output_dense_bias_to_fp16 = const()[name = string("layers_3_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28098752)))];
+            tensor<fp16, [1, 128, 384]> linear_21_cast_fp16 = linear(bias = layers_3_attention_output_dense_bias_to_fp16, weight = layers_3_attention_output_dense_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("linear_21_cast_fp16")];
+            tensor<fp16, [1, 128, 384]> input_79_cast_fp16 = add(x = linear_21_cast_fp16, y = input_71_cast_fp16)[name = string("input_79_cast_fp16")];
+            tensor<int32, [1]> input_81_axes_0 = const()[name = string("input_81_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_3_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_3_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28099584)))];
+            tensor<fp16, [384]> layers_3_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_3_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28100416)))];
+            fp16 var_622_to_fp16 = const()[name = string("op_622_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 128, 384]> input_81_cast_fp16 = layer_norm(axes = input_81_axes_0, beta = layers_3_attention_output_LayerNorm_bias_to_fp16, epsilon = var_622_to_fp16, gamma = layers_3_attention_output_LayerNorm_weight_to_fp16, x = input_79_cast_fp16)[name = string("input_81_cast_fp16")];
+            tensor<fp16, [1536, 384]> layers_3_intermediate_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28101248))), lut = tensor<fp16, [96, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28543680))))[name = string("layers_3_intermediate_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [1536]> layers_3_intermediate_dense_bias_to_fp16 = const()[name = string("layers_3_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28556032)))];
+            tensor<fp16, [1, 128, 1536]> linear_22_cast_fp16 = linear(bias = layers_3_intermediate_dense_bias_to_fp16, weight = layers_3_intermediate_dense_weight_to_fp16_palettized, x = input_81_cast_fp16)[name = string("linear_22_cast_fp16")];
+            string input_85_mode_0 = const()[name = string("input_85_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 128, 1536]> input_85_cast_fp16 = gelu(mode = input_85_mode_0, x = linear_22_cast_fp16)[name = string("input_85_cast_fp16")];
+            tensor<fp16, [384, 1536]> layers_3_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28559168))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29001600))))[name = string("layers_3_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_3_output_dense_bias_to_fp16 = const()[name = string("layers_3_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29004736)))];
+            tensor<fp16, [1, 128, 384]> linear_23_cast_fp16 = linear(bias = layers_3_output_dense_bias_to_fp16, weight = layers_3_output_dense_weight_to_fp16_palettized, x = input_85_cast_fp16)[name = string("linear_23_cast_fp16")];
+            tensor<fp16, [1, 128, 384]> input_89_cast_fp16 = add(x = linear_23_cast_fp16, y = input_81_cast_fp16)[name = string("input_89_cast_fp16")];
+            tensor<int32, [1]> input_91_axes_0 = const()[name = string("input_91_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_3_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_3_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29005568)))];
+            tensor<fp16, [384]> layers_3_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_3_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29006400)))];
+            fp16 var_645_to_fp16 = const()[name = string("op_645_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 128, 384]> input_91_cast_fp16 = layer_norm(axes = input_91_axes_0, beta = layers_3_output_LayerNorm_bias_to_fp16, epsilon = var_645_to_fp16, gamma = layers_3_output_LayerNorm_weight_to_fp16, x = input_89_cast_fp16)[name = string("input_91_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_4_attention_self_query_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29007232))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29117888))))[name = string("layers_4_attention_self_query_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_4_attention_self_query_bias_to_fp16 = const()[name = string("layers_4_attention_self_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29121024)))];
+            tensor<fp16, [1, 128, 384]> linear_24_cast_fp16 = linear(bias = layers_4_attention_self_query_bias_to_fp16, weight = layers_4_attention_self_query_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = string("linear_24_cast_fp16")];
+            tensor<int32, [4]> var_668 = const()[name = string("op_668"), val = tensor<int32, [4]>([1, 128, 12, 32])];
+            tensor<fp16, [1, 128, 12, 32]> var_669_cast_fp16 = reshape(shape = var_668, x = linear_24_cast_fp16)[name = string("op_669_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_4_attention_self_key_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29121856))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29232512))))[name = string("layers_4_attention_self_key_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_4_attention_self_key_bias_to_fp16 = const()[name = string("layers_4_attention_self_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29235648)))];
+            tensor<fp16, [1, 128, 384]> linear_25_cast_fp16 = linear(bias = layers_4_attention_self_key_bias_to_fp16, weight = layers_4_attention_self_key_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = string("linear_25_cast_fp16")];
+            tensor<int32, [4]> var_683 = const()[name = string("op_683"), val = tensor<int32, [4]>([1, 128, 12, 32])];
+            tensor<fp16, [1, 128, 12, 32]> var_684_cast_fp16 = reshape(shape = var_683, x = linear_25_cast_fp16)[name = string("op_684_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_4_attention_self_value_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29236480))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29347136))))[name = string("layers_4_attention_self_value_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_4_attention_self_value_bias_to_fp16 = const()[name = string("layers_4_attention_self_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29350272)))];
+            tensor<fp16, [1, 128, 384]> linear_26_cast_fp16 = linear(bias = layers_4_attention_self_value_bias_to_fp16, weight = layers_4_attention_self_value_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = string("linear_26_cast_fp16")];
+            tensor<int32, [4]> var_698 = const()[name = string("op_698"), val = tensor<int32, [4]>([1, 128, 12, 32])];
+            tensor<fp16, [1, 128, 12, 32]> var_699_cast_fp16 = reshape(shape = var_698, x = linear_26_cast_fp16)[name = string("op_699_cast_fp16")];
+            tensor<int32, [4]> var_704 = const()[name = string("op_704"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool var_709_transpose_x_0 = const()[name = string("op_709_transpose_x_0"), val = bool(false)];
+            bool var_709_transpose_y_0 = const()[name = string("op_709_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 12, 32, 128]> transpose_33 = transpose(perm = transpose_33_perm_0, x = var_684_cast_fp16)[name = string("transpose_42")];
+            tensor<fp16, [1, 12, 128, 32]> transpose_32 = transpose(perm = transpose_32_perm_0, x = var_669_cast_fp16)[name = string("transpose_43")];
+            tensor<fp16, [1, 12, 128, 128]> var_709_cast_fp16 = matmul(transpose_x = var_709_transpose_x_0, transpose_y = var_709_transpose_y_0, x = transpose_32, y = transpose_33)[name = string("op_709_cast_fp16")];
+            fp16 var_710_to_fp16 = const()[name = string("op_710_to_fp16"), val = fp16(0x1.6ap-3)];
+            tensor<fp16, [1, 12, 128, 128]> weights_17_cast_fp16 = mul(x = var_709_cast_fp16, y = var_710_to_fp16)[name = string("weights_17_cast_fp16")];
+            tensor<fp16, [1, 12, 128, 128]> var_713_cast_fp16 = add(x = weights_17_cast_fp16, y = additive_mask_cast_fp16)[name = string("op_713_cast_fp16")];
+            int32 var_714 = const()[name = string("op_714"), val = int32(-1)];
+            tensor<fp16, [1, 12, 128, 128]> input_93_cast_fp16 = softmax(axis = var_714, x = var_713_cast_fp16)[name = string("input_93_cast_fp16")];
+            bool var_720_transpose_x_0 = const()[name = string("op_720_transpose_x_0"), val = bool(false)];
+            bool var_720_transpose_y_0 = const()[name = string("op_720_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 128, 32]> value_9_cast_fp16 = transpose(perm = var_704, x = var_699_cast_fp16)[name = string("transpose_41")];
+            tensor<fp16, [1, 12, 128, 32]> var_720_cast_fp16 = matmul(transpose_x = var_720_transpose_x_0, transpose_y = var_720_transpose_y_0, x = input_93_cast_fp16, y = value_9_cast_fp16)[name = string("op_720_cast_fp16")];
+            tensor<int32, [4]> var_725 = const()[name = string("op_725"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_730 = const()[name = string("op_730"), val = tensor<int32, [3]>([1, 128, 384])];
+            tensor<fp16, [1, 128, 12, 32]> var_726_cast_fp16 = transpose(perm = var_725, x = var_720_cast_fp16)[name = string("transpose_40")];
+            tensor<fp16, [1, 128, 384]> input_95_cast_fp16 = reshape(shape = var_730, x = var_726_cast_fp16)[name = string("input_95_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_4_attention_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29351104))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29461760))))[name = string("layers_4_attention_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_4_attention_output_dense_bias_to_fp16 = const()[name = string("layers_4_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29464896)))];
+            tensor<fp16, [1, 128, 384]> linear_27_cast_fp16 = linear(bias = layers_4_attention_output_dense_bias_to_fp16, weight = layers_4_attention_output_dense_weight_to_fp16_palettized, x = input_95_cast_fp16)[name = string("linear_27_cast_fp16")];
+            tensor<fp16, [1, 128, 384]> input_99_cast_fp16 = add(x = linear_27_cast_fp16, y = input_91_cast_fp16)[name = string("input_99_cast_fp16")];
+            tensor<int32, [1]> input_101_axes_0 = const()[name = string("input_101_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_4_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_4_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29465728)))];
+            tensor<fp16, [384]> layers_4_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_4_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29466560)))];
+            fp16 var_733_to_fp16 = const()[name = string("op_733_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 128, 384]> input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = layers_4_attention_output_LayerNorm_bias_to_fp16, epsilon = var_733_to_fp16, gamma = layers_4_attention_output_LayerNorm_weight_to_fp16, x = input_99_cast_fp16)[name = string("input_101_cast_fp16")];
+            tensor<fp16, [1536, 384]> layers_4_intermediate_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29467392))), lut = tensor<fp16, [96, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29909824))))[name = string("layers_4_intermediate_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [1536]> layers_4_intermediate_dense_bias_to_fp16 = const()[name = string("layers_4_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29922176)))];
+            tensor<fp16, [1, 128, 1536]> linear_28_cast_fp16 = linear(bias = layers_4_intermediate_dense_bias_to_fp16, weight = layers_4_intermediate_dense_weight_to_fp16_palettized, x = input_101_cast_fp16)[name = string("linear_28_cast_fp16")];
+            string input_105_mode_0 = const()[name = string("input_105_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 128, 1536]> input_105_cast_fp16 = gelu(mode = input_105_mode_0, x = linear_28_cast_fp16)[name = string("input_105_cast_fp16")];
+            tensor<fp16, [384, 1536]> layers_4_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29925312))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30367744))))[name = string("layers_4_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_4_output_dense_bias_to_fp16 = const()[name = string("layers_4_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30370880)))];
+            tensor<fp16, [1, 128, 384]> linear_29_cast_fp16 = linear(bias = layers_4_output_dense_bias_to_fp16, weight = layers_4_output_dense_weight_to_fp16_palettized, x = input_105_cast_fp16)[name = string("linear_29_cast_fp16")];
+            tensor<fp16, [1, 128, 384]> input_109_cast_fp16 = add(x = linear_29_cast_fp16, y = input_101_cast_fp16)[name = string("input_109_cast_fp16")];
+            tensor<int32, [1]> input_111_axes_0 = const()[name = string("input_111_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_4_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_4_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30371712)))];
+            tensor<fp16, [384]> layers_4_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_4_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30372544)))];
+            fp16 var_756_to_fp16 = const()[name = string("op_756_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 128, 384]> input_111_cast_fp16 = layer_norm(axes = input_111_axes_0, beta = layers_4_output_LayerNorm_bias_to_fp16, epsilon = var_756_to_fp16, gamma = layers_4_output_LayerNorm_weight_to_fp16, x = input_109_cast_fp16)[name = string("input_111_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_5_attention_self_query_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30373376))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30484032))))[name = string("layers_5_attention_self_query_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_5_attention_self_query_bias_to_fp16 = const()[name = string("layers_5_attention_self_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30487168)))];
+            tensor<fp16, [1, 128, 384]> linear_30_cast_fp16 = linear(bias = layers_5_attention_self_query_bias_to_fp16, weight = layers_5_attention_self_query_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = string("linear_30_cast_fp16")];
+            tensor<int32, [4]> var_779 = const()[name = string("op_779"), val = tensor<int32, [4]>([1, 128, 12, 32])];
+            tensor<fp16, [1, 128, 12, 32]> var_780_cast_fp16 = reshape(shape = var_779, x = linear_30_cast_fp16)[name = string("op_780_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_5_attention_self_key_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30488000))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30598656))))[name = string("layers_5_attention_self_key_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_5_attention_self_key_bias_to_fp16 = const()[name = string("layers_5_attention_self_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30601792)))];
+            tensor<fp16, [1, 128, 384]> linear_31_cast_fp16 = linear(bias = layers_5_attention_self_key_bias_to_fp16, weight = layers_5_attention_self_key_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = string("linear_31_cast_fp16")];
+            tensor<int32, [4]> var_794 = const()[name = string("op_794"), val = tensor<int32, [4]>([1, 128, 12, 32])];
+            tensor<fp16, [1, 128, 12, 32]> var_795_cast_fp16 = reshape(shape = var_794, x = linear_31_cast_fp16)[name = string("op_795_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_5_attention_self_value_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30602624))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30713280))))[name = string("layers_5_attention_self_value_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_5_attention_self_value_bias_to_fp16 = const()[name = string("layers_5_attention_self_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30716416)))];
+            tensor<fp16, [1, 128, 384]> linear_32_cast_fp16 = linear(bias = layers_5_attention_self_value_bias_to_fp16, weight = layers_5_attention_self_value_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = string("linear_32_cast_fp16")];
+            tensor<int32, [4]> var_809 = const()[name = string("op_809"), val = tensor<int32, [4]>([1, 128, 12, 32])];
+            tensor<fp16, [1, 128, 12, 32]> var_810_cast_fp16 = reshape(shape = var_809, x = linear_32_cast_fp16)[name = string("op_810_cast_fp16")];
+            tensor<int32, [4]> var_815 = const()[name = string("op_815"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool var_820_transpose_x_0 = const()[name = string("op_820_transpose_x_0"), val = bool(false)];
+            bool var_820_transpose_y_0 = const()[name = string("op_820_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 12, 32, 128]> transpose_35 = transpose(perm = transpose_35_perm_0, x = var_795_cast_fp16)[name = string("transpose_38")];
+            tensor<fp16, [1, 12, 128, 32]> transpose_34 = transpose(perm = transpose_34_perm_0, x = var_780_cast_fp16)[name = string("transpose_39")];
+            tensor<fp16, [1, 12, 128, 128]> var_820_cast_fp16 = matmul(transpose_x = var_820_transpose_x_0, transpose_y = var_820_transpose_y_0, x = transpose_34, y = transpose_35)[name = string("op_820_cast_fp16")];
+            fp16 var_821_to_fp16 = const()[name = string("op_821_to_fp16"), val = fp16(0x1.6ap-3)];
+            tensor<fp16, [1, 12, 128, 128]> weights_21_cast_fp16 = mul(x = var_820_cast_fp16, y = var_821_to_fp16)[name = string("weights_21_cast_fp16")];
+            tensor<fp16, [1, 12, 128, 128]> var_824_cast_fp16 = add(x = weights_21_cast_fp16, y = additive_mask_cast_fp16)[name = string("op_824_cast_fp16")];
+            int32 var_825 = const()[name = string("op_825"), val = int32(-1)];
+            tensor<fp16, [1, 12, 128, 128]> input_113_cast_fp16 = softmax(axis = var_825, x = var_824_cast_fp16)[name = string("input_113_cast_fp16")];
+            bool var_831_transpose_x_0 = const()[name = string("op_831_transpose_x_0"), val = bool(false)];
+            bool var_831_transpose_y_0 = const()[name = string("op_831_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 128, 32]> value_11_cast_fp16 = transpose(perm = var_815, x = var_810_cast_fp16)[name = string("transpose_37")];
+            tensor<fp16, [1, 12, 128, 32]> var_831_cast_fp16 = matmul(transpose_x = var_831_transpose_x_0, transpose_y = var_831_transpose_y_0, x = input_113_cast_fp16, y = value_11_cast_fp16)[name = string("op_831_cast_fp16")];
+            tensor<int32, [4]> var_836 = const()[name = string("op_836"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_841 = const()[name = string("op_841"), val = tensor<int32, [3]>([1, 128, 384])];
+            tensor<fp16, [1, 128, 12, 32]> var_837_cast_fp16 = transpose(perm = var_836, x = var_831_cast_fp16)[name = string("transpose_36")];
+            tensor<fp16, [1, 128, 384]> input_115_cast_fp16 = reshape(shape = var_841, x = var_837_cast_fp16)[name = string("input_115_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_5_attention_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30717248))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30827904))))[name = string("layers_5_attention_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_5_attention_output_dense_bias_to_fp16 = const()[name = string("layers_5_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30831040)))];
+            tensor<fp16, [1, 128, 384]> linear_33_cast_fp16 = linear(bias = layers_5_attention_output_dense_bias_to_fp16, weight = layers_5_attention_output_dense_weight_to_fp16_palettized, x = input_115_cast_fp16)[name = string("linear_33_cast_fp16")];
+            tensor<fp16, [1, 128, 384]> input_119_cast_fp16 = add(x = linear_33_cast_fp16, y = input_111_cast_fp16)[name = string("input_119_cast_fp16")];
+            tensor<int32, [1]> input_121_axes_0 = const()[name = string("input_121_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_5_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_5_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30831872)))];
+            tensor<fp16, [384]> layers_5_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_5_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30832704)))];
+            fp16 var_844_to_fp16 = const()[name = string("op_844_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 128, 384]> input_121_cast_fp16 = layer_norm(axes = input_121_axes_0, beta = layers_5_attention_output_LayerNorm_bias_to_fp16, epsilon = var_844_to_fp16, gamma = layers_5_attention_output_LayerNorm_weight_to_fp16, x = input_119_cast_fp16)[name = string("input_121_cast_fp16")];
+            tensor<fp16, [1536, 384]> layers_5_intermediate_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30833536))), lut = tensor<fp16, [96, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31275968))))[name = string("layers_5_intermediate_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [1536]> layers_5_intermediate_dense_bias_to_fp16 = const()[name = string("layers_5_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31288320)))];
+            tensor<fp16, [1, 128, 1536]> linear_34_cast_fp16 = linear(bias = layers_5_intermediate_dense_bias_to_fp16, weight = layers_5_intermediate_dense_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("linear_34_cast_fp16")];
+            string input_125_mode_0 = const()[name = string("input_125_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 128, 1536]> input_125_cast_fp16 = gelu(mode = input_125_mode_0, x = linear_34_cast_fp16)[name = string("input_125_cast_fp16")];
+            tensor<fp16, [384, 1536]> layers_5_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31291456))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31733888))))[name = string("layers_5_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_5_output_dense_bias_to_fp16 = const()[name = string("layers_5_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31737024)))];
+            tensor<fp16, [1, 128, 384]> linear_35_cast_fp16 = linear(bias = layers_5_output_dense_bias_to_fp16, weight = layers_5_output_dense_weight_to_fp16_palettized, x = input_125_cast_fp16)[name = string("linear_35_cast_fp16")];
+            tensor<fp16, [1, 128, 384]> input_129_cast_fp16 = add(x = linear_35_cast_fp16, y = input_121_cast_fp16)[name = string("input_129_cast_fp16")];
+            tensor<int32, [1]> hidden_states_axes_0 = const()[name = string("hidden_states_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_5_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_5_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31737856)))];
+            tensor<fp16, [384]> layers_5_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_5_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31738688)))];
+            fp16 var_867_to_fp16 = const()[name = string("op_867_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 128, 384]> hidden_states_cast_fp16 = layer_norm(axes = hidden_states_axes_0, beta = layers_5_output_LayerNorm_bias_to_fp16, epsilon = var_867_to_fp16, gamma = layers_5_output_LayerNorm_weight_to_fp16, x = input_129_cast_fp16)[name = string("hidden_states_cast_fp16")];
+            tensor<int32, [3]> input_131_begin_0 = const()[name = string("input_131_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> input_131_end_0 = const()[name = string("input_131_end_0"), val = tensor<int32, [3]>([1, 1, 384])];
+            tensor<bool, [3]> input_131_end_mask_0 = const()[name = string("input_131_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<bool, [3]> input_131_squeeze_mask_0 = const()[name = string("input_131_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
+            tensor<fp16, [1, 384]> input_131_cast_fp16 = slice_by_index(begin = input_131_begin_0, end = input_131_end_0, end_mask = input_131_end_mask_0, squeeze_mask = input_131_squeeze_mask_0, x = hidden_states_cast_fp16)[name = string("input_131_cast_fp16")];
+            tensor<fp16, [384, 384]> pooler_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31739520))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31850176))))[name = string("pooler_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> pooler_dense_bias_to_fp16 = const()[name = string("pooler_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31853312)))];
+            tensor<fp16, [1, 384]> linear_36_cast_fp16 = linear(bias = pooler_dense_bias_to_fp16, weight = pooler_dense_weight_to_fp16_palettized, x = input_131_cast_fp16)[name = string("linear_36_cast_fp16")];
+            tensor<fp16, [1, 384]> input_135_cast_fp16 = tanh(x = linear_36_cast_fp16)[name = string("input_135_cast_fp16")];
+            tensor<fp16, [1, 384]> classifier_weight_to_fp16 = const()[name = string("classifier_weight_to_fp16"), val = tensor<fp16, [1, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31854144)))];
+            tensor<fp16, [1]> classifier_bias_to_fp16 = const()[name = string("classifier_bias_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
+            tensor<fp16, [1, 1]> linear_37_cast_fp16 = linear(bias = classifier_bias_to_fp16, weight = classifier_weight_to_fp16, x = input_135_cast_fp16)[name = string("linear_37_cast_fp16")];
+            string linear_37_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_37_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 1]> score = cast(dtype = linear_37_cast_fp16_to_fp32_dtype_0, x = linear_37_cast_fp16)[name = string("cast_1")];
+        } -> (score);
+}

minilm_reranker_128.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:037e8710fad94dc991a1fe3ecaf8f125883ac60d3fad6b01c6b6c346ecd86430
+size 31854976

minilm_reranker_256.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cea7713269e7cdc358452a7ef2430c0eae940720c403e17c78397be9de555d13
+size 243

minilm_reranker_256.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5fbc7ba55d6ef9035268e5da78df5dac91a5caf89b5b0f7297c18f33556ac96b
+size 434

minilm_reranker_256.mlmodelc/metadata.json

@@ -0,0 +1,97 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Mixed (Float16, Palettized (11 bits), Palettized (13 bits), UInt6)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1]",
+        "name" : "score",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 9,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios18.linear" : 38,
+      "Ios18.constexprLutToDense" : 37,
+      "Ios18.sub" : 1,
+      "Select" : 2,
+      "Ios18.matmul" : 12,
+      "Ios18.gelu" : 6,
+      "Ios18.gather" : 2,
+      "Ios18.add" : 22,
+      "Ios18.tanh" : 1,
+      "Ios18.softmax" : 6,
+      "Ios18.sliceByIndex" : 1,
+      "Ios18.layerNorm" : 13,
+      "Ios18.transpose" : 24,
+      "Ios18.cast" : 7,
+      "Ios18.reshape" : 25,
+      "Ios18.greaterEqual" : 2,
+      "Ios18.mul" : 7
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int16, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "15.0",
+      "tvOS" : "18.0",
+      "visionOS" : "2.0",
+      "watchOS" : "11.0",
+      "iOS" : "18.0",
+      "macCatalyst" : "18.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-07-15",
+      "com.github.apple.coremltools.source" : "torch==2.5.0",
+      "com.github.apple.coremltools.version" : "9.0",
+      "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"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 256)",
+        "shortDescription" : "",
+        "shape" : "[1, 256]",
+        "name" : "attention_mask",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 256)",
+        "shortDescription" : "",
+        "shape" : "[1, 256]",
+        "name" : "token_type_ids",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "minilm_reranker_256",
+    "method" : "predict"
+  }
+]

minilm_reranker_256.mlmodelc/model.mil

@@ -0,0 +1,420 @@
+program(1.3)
+[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}})]
+{
+    func main<ios18>(tensor<int32, [1, 256]> attention_mask, tensor<int32, [1, 256]> input_ids, tensor<int32, [1, 256]> token_type_ids) {
+            int32 hidden_1_batch_dims_0 = const()[name = string("hidden_1_batch_dims_0"), val = int32(0)];
+            bool hidden_1_validate_indices_0 = const()[name = string("hidden_1_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [30522, 384]> embeddings_word_embeddings_weight_to_fp16 = const()[name = string("embeddings_word_embeddings_weight_to_fp16"), val = tensor<fp16, [30522, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
+            string input_ids_to_int16_dtype_0 = const()[name = string("input_ids_to_int16_dtype_0"), val = string("int16")];
+            string cast_28_dtype_0 = const()[name = string("cast_28_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
+            tensor<int16, [1, 256]> input_ids_to_int16 = cast(dtype = input_ids_to_int16_dtype_0, x = input_ids)[name = string("cast_7")];
+            tensor<int32, [1, 256]> cast_28 = cast(dtype = cast_28_dtype_0, x = input_ids_to_int16)[name = string("cast_6")];
+            tensor<bool, [1, 256]> greater_equal_0 = greater_equal(x = cast_28, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
+            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(30522)];
+            tensor<int32, [1, 256]> add_0 = add(x = cast_28, y = slice_by_index_0)[name = string("add_0")];
+            tensor<int32, [1, 256]> select_0 = select(a = cast_28, b = add_0, cond = greater_equal_0)[name = string("select_0")];
+            string select_0_to_int16_dtype_0 = const()[name = string("select_0_to_int16_dtype_0"), val = string("int16")];
+            string cast_0_dtype_0 = const()[name = string("cast_0_dtype_0"), val = string("int32")];
+            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
+            tensor<int16, [1, 256]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = string("cast_5")];
+            tensor<int32, [1, 256]> cast_0 = cast(dtype = cast_0_dtype_0, x = select_0_to_int16)[name = string("cast_4")];
+            tensor<bool, [1, 256]> greater_equal_0_1 = greater_equal(x = cast_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
+            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(30522)];
+            tensor<int32, [1, 256]> add_0_1 = add(x = cast_0, y = slice_by_index_0_1)[name = string("add_0_1")];
+            tensor<int32, [1, 256]> select_0_1 = select(a = cast_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")];
+            int32 hidden_1_cast_fp16_cast_uint16_cast_uint16_axis_0 = const()[name = string("hidden_1_cast_fp16_cast_uint16_cast_uint16_axis_0"), val = int32(0)];
+            tensor<fp16, [1, 256, 384]> hidden_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = hidden_1_cast_fp16_cast_uint16_cast_uint16_axis_0, batch_dims = hidden_1_batch_dims_0, indices = select_0_1, validate_indices = hidden_1_validate_indices_0, x = embeddings_word_embeddings_weight_to_fp16)[name = string("hidden_1_cast_fp16_cast_uint16_cast_uint16")];
+            int32 var_182_axis_0 = const()[name = string("op_182_axis_0"), val = int32(0)];
+            int32 var_182_batch_dims_0 = const()[name = string("op_182_batch_dims_0"), val = int32(0)];
+            bool var_182_validate_indices_0 = const()[name = string("op_182_validate_indices_0"), val = bool(false)];
+            tensor<fp16, [2, 384]> embeddings_token_type_embeddings_weight_to_fp16 = const()[name = string("embeddings_token_type_embeddings_weight_to_fp16"), val = tensor<fp16, [2, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23441024)))];
+            string token_type_ids_to_uint16_dtype_0 = const()[name = string("token_type_ids_to_uint16_dtype_0"), val = string("uint16")];
+            tensor<uint16, [1, 256]> token_type_ids_to_uint16 = cast(dtype = token_type_ids_to_uint16_dtype_0, x = token_type_ids)[name = string("cast_3")];
+            tensor<fp16, [1, 256, 384]> var_182_cast_fp16_cast_uint16 = gather(axis = var_182_axis_0, batch_dims = var_182_batch_dims_0, indices = token_type_ids_to_uint16, validate_indices = var_182_validate_indices_0, x = embeddings_token_type_embeddings_weight_to_fp16)[name = string("op_182_cast_fp16_cast_uint16")];
+            tensor<fp16, [1, 256, 384]> hidden_cast_fp16 = add(x = hidden_1_cast_fp16_cast_uint16_cast_uint16, y = var_182_cast_fp16_cast_uint16)[name = string("hidden_cast_fp16")];
+            tensor<fp16, [1, 256, 384]> var_188_to_fp16 = const()[name = string("op_188_to_fp16"), val = tensor<fp16, [1, 256, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23442624)))];
+            tensor<fp16, [1, 256, 384]> input_7_cast_fp16 = add(x = hidden_cast_fp16, y = var_188_to_fp16)[name = string("input_7_cast_fp16")];
+            tensor<int32, [1]> input_9_axes_0 = const()[name = string("input_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> embeddings_LayerNorm_weight_to_fp16 = const()[name = string("embeddings_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23639296)))];
+            tensor<fp16, [384]> embeddings_LayerNorm_bias_to_fp16 = const()[name = string("embeddings_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23640128)))];
+            fp16 var_192_to_fp16 = const()[name = string("op_192_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 256, 384]> input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = embeddings_LayerNorm_bias_to_fp16, epsilon = var_192_to_fp16, gamma = embeddings_LayerNorm_weight_to_fp16, x = input_7_cast_fp16)[name = string("input_9_cast_fp16")];
+            fp16 var_206_to_fp16 = const()[name = string("op_206_to_fp16"), val = fp16(0x1p+0)];
+            string var_205_to_fp16_dtype_0 = const()[name = string("op_205_to_fp16_dtype_0"), val = string("fp16")];
+            tensor<fp16, [1, 256]> attention_mask_to_fp16 = cast(dtype = var_205_to_fp16_dtype_0, x = attention_mask)[name = string("cast_2")];
+            tensor<fp16, [1, 256]> var_208_cast_fp16 = sub(x = var_206_to_fp16, y = attention_mask_to_fp16)[name = string("op_208_cast_fp16")];
+            tensor<int32, [4]> var_213 = const()[name = string("op_213"), val = tensor<int32, [4]>([1, 1, 1, 256])];
+            tensor<fp16, [1, 1, 1, 256]> var_214_cast_fp16 = reshape(shape = var_213, x = var_208_cast_fp16)[name = string("op_214_cast_fp16")];
+            fp16 var_215_to_fp16 = const()[name = string("op_215_to_fp16"), val = fp16(-0x1.388p+13)];
+            tensor<fp16, [1, 1, 1, 256]> additive_mask_cast_fp16 = mul(x = var_214_cast_fp16, y = var_215_to_fp16)[name = string("additive_mask_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_0_attention_self_query_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23640960))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23751616))))[name = string("layers_0_attention_self_query_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_0_attention_self_query_bias_to_fp16 = const()[name = string("layers_0_attention_self_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23754752)))];
+            tensor<fp16, [1, 256, 384]> linear_0_cast_fp16 = linear(bias = layers_0_attention_self_query_bias_to_fp16, weight = layers_0_attention_self_query_weight_to_fp16_palettized, x = input_9_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [4]> var_224 = const()[name = string("op_224"), val = tensor<int32, [4]>([1, 256, 12, 32])];
+            tensor<fp16, [1, 256, 12, 32]> var_225_cast_fp16 = reshape(shape = var_224, x = linear_0_cast_fp16)[name = string("op_225_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_0_attention_self_key_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23755584))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23866240))))[name = string("layers_0_attention_self_key_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_0_attention_self_key_bias_to_fp16 = const()[name = string("layers_0_attention_self_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23869376)))];
+            tensor<fp16, [1, 256, 384]> linear_1_cast_fp16 = linear(bias = layers_0_attention_self_key_bias_to_fp16, weight = layers_0_attention_self_key_weight_to_fp16_palettized, x = input_9_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<int32, [4]> var_239 = const()[name = string("op_239"), val = tensor<int32, [4]>([1, 256, 12, 32])];
+            tensor<fp16, [1, 256, 12, 32]> var_240_cast_fp16 = reshape(shape = var_239, x = linear_1_cast_fp16)[name = string("op_240_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_0_attention_self_value_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23870208))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23980864))))[name = string("layers_0_attention_self_value_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_0_attention_self_value_bias_to_fp16 = const()[name = string("layers_0_attention_self_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23984000)))];
+            tensor<fp16, [1, 256, 384]> linear_2_cast_fp16 = linear(bias = layers_0_attention_self_value_bias_to_fp16, weight = layers_0_attention_self_value_weight_to_fp16_palettized, x = input_9_cast_fp16)[name = string("linear_2_cast_fp16")];
+            tensor<int32, [4]> var_254 = const()[name = string("op_254"), val = tensor<int32, [4]>([1, 256, 12, 32])];
+            tensor<fp16, [1, 256, 12, 32]> var_255_cast_fp16 = reshape(shape = var_254, x = linear_2_cast_fp16)[name = string("op_255_cast_fp16")];
+            tensor<int32, [4]> var_260 = const()[name = string("op_260"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool var_265_transpose_x_0 = const()[name = string("op_265_transpose_x_0"), val = bool(false)];
+            bool var_265_transpose_y_0 = const()[name = string("op_265_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_24_perm_0 = const()[name = string("transpose_24_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_25_perm_0 = const()[name = string("transpose_25_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 12, 32, 256]> transpose_25 = transpose(perm = transpose_25_perm_0, x = var_240_cast_fp16)[name = string("transpose_58")];
+            tensor<fp16, [1, 12, 256, 32]> transpose_24 = transpose(perm = transpose_24_perm_0, x = var_225_cast_fp16)[name = string("transpose_59")];
+            tensor<fp16, [1, 12, 256, 256]> var_265_cast_fp16 = matmul(transpose_x = var_265_transpose_x_0, transpose_y = var_265_transpose_y_0, x = transpose_24, y = transpose_25)[name = string("op_265_cast_fp16")];
+            fp16 var_266_to_fp16 = const()[name = string("op_266_to_fp16"), val = fp16(0x1.6ap-3)];
+            tensor<fp16, [1, 12, 256, 256]> weights_1_cast_fp16 = mul(x = var_265_cast_fp16, y = var_266_to_fp16)[name = string("weights_1_cast_fp16")];
+            tensor<fp16, [1, 12, 256, 256]> var_269_cast_fp16 = add(x = weights_1_cast_fp16, y = additive_mask_cast_fp16)[name = string("op_269_cast_fp16")];
+            int32 var_270 = const()[name = string("op_270"), val = int32(-1)];
+            tensor<fp16, [1, 12, 256, 256]> input_13_cast_fp16 = softmax(axis = var_270, x = var_269_cast_fp16)[name = string("input_13_cast_fp16")];
+            bool var_276_transpose_x_0 = const()[name = string("op_276_transpose_x_0"), val = bool(false)];
+            bool var_276_transpose_y_0 = const()[name = string("op_276_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 256, 32]> value_1_cast_fp16 = transpose(perm = var_260, x = var_255_cast_fp16)[name = string("transpose_57")];
+            tensor<fp16, [1, 12, 256, 32]> var_276_cast_fp16 = matmul(transpose_x = var_276_transpose_x_0, transpose_y = var_276_transpose_y_0, x = input_13_cast_fp16, y = value_1_cast_fp16)[name = string("op_276_cast_fp16")];
+            tensor<int32, [4]> var_281 = const()[name = string("op_281"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_286 = const()[name = string("op_286"), val = tensor<int32, [3]>([1, 256, 384])];
+            tensor<fp16, [1, 256, 12, 32]> var_282_cast_fp16 = transpose(perm = var_281, x = var_276_cast_fp16)[name = string("transpose_56")];
+            tensor<fp16, [1, 256, 384]> input_15_cast_fp16 = reshape(shape = var_286, x = var_282_cast_fp16)[name = string("input_15_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_0_attention_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23984832))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24095488))))[name = string("layers_0_attention_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_0_attention_output_dense_bias_to_fp16 = const()[name = string("layers_0_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24098624)))];
+            tensor<fp16, [1, 256, 384]> linear_3_cast_fp16 = linear(bias = layers_0_attention_output_dense_bias_to_fp16, weight = layers_0_attention_output_dense_weight_to_fp16_palettized, x = input_15_cast_fp16)[name = string("linear_3_cast_fp16")];
+            tensor<fp16, [1, 256, 384]> input_19_cast_fp16 = add(x = linear_3_cast_fp16, y = input_9_cast_fp16)[name = string("input_19_cast_fp16")];
+            tensor<int32, [1]> input_21_axes_0 = const()[name = string("input_21_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_0_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_0_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24099456)))];
+            tensor<fp16, [384]> layers_0_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_0_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24100288)))];
+            fp16 var_289_to_fp16 = const()[name = string("op_289_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 256, 384]> input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = layers_0_attention_output_LayerNorm_bias_to_fp16, epsilon = var_289_to_fp16, gamma = layers_0_attention_output_LayerNorm_weight_to_fp16, x = input_19_cast_fp16)[name = string("input_21_cast_fp16")];
+            tensor<fp16, [1536, 384]> layers_0_intermediate_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24101120))), lut = tensor<fp16, [96, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24543552))))[name = string("layers_0_intermediate_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [1536]> layers_0_intermediate_dense_bias_to_fp16 = const()[name = string("layers_0_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24555904)))];
+            tensor<fp16, [1, 256, 1536]> linear_4_cast_fp16 = linear(bias = layers_0_intermediate_dense_bias_to_fp16, weight = layers_0_intermediate_dense_weight_to_fp16_palettized, x = input_21_cast_fp16)[name = string("linear_4_cast_fp16")];
+            string input_25_mode_0 = const()[name = string("input_25_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 256, 1536]> input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_4_cast_fp16)[name = string("input_25_cast_fp16")];
+            tensor<fp16, [384, 1536]> layers_0_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24559040))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25001472))))[name = string("layers_0_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_0_output_dense_bias_to_fp16 = const()[name = string("layers_0_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25004608)))];
+            tensor<fp16, [1, 256, 384]> linear_5_cast_fp16 = linear(bias = layers_0_output_dense_bias_to_fp16, weight = layers_0_output_dense_weight_to_fp16_palettized, x = input_25_cast_fp16)[name = string("linear_5_cast_fp16")];
+            tensor<fp16, [1, 256, 384]> input_29_cast_fp16 = add(x = linear_5_cast_fp16, y = input_21_cast_fp16)[name = string("input_29_cast_fp16")];
+            tensor<int32, [1]> input_31_axes_0 = const()[name = string("input_31_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_0_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_0_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25005440)))];
+            tensor<fp16, [384]> layers_0_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_0_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25006272)))];
+            fp16 var_312_to_fp16 = const()[name = string("op_312_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 256, 384]> input_31_cast_fp16 = layer_norm(axes = input_31_axes_0, beta = layers_0_output_LayerNorm_bias_to_fp16, epsilon = var_312_to_fp16, gamma = layers_0_output_LayerNorm_weight_to_fp16, x = input_29_cast_fp16)[name = string("input_31_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_1_attention_self_query_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25007104))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25117760))))[name = string("layers_1_attention_self_query_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_1_attention_self_query_bias_to_fp16 = const()[name = string("layers_1_attention_self_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25120896)))];
+            tensor<fp16, [1, 256, 384]> linear_6_cast_fp16 = linear(bias = layers_1_attention_self_query_bias_to_fp16, weight = layers_1_attention_self_query_weight_to_fp16_palettized, x = input_31_cast_fp16)[name = string("linear_6_cast_fp16")];
+            tensor<int32, [4]> var_335 = const()[name = string("op_335"), val = tensor<int32, [4]>([1, 256, 12, 32])];
+            tensor<fp16, [1, 256, 12, 32]> var_336_cast_fp16 = reshape(shape = var_335, x = linear_6_cast_fp16)[name = string("op_336_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_1_attention_self_key_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25121728))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25232384))))[name = string("layers_1_attention_self_key_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_1_attention_self_key_bias_to_fp16 = const()[name = string("layers_1_attention_self_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25235520)))];
+            tensor<fp16, [1, 256, 384]> linear_7_cast_fp16 = linear(bias = layers_1_attention_self_key_bias_to_fp16, weight = layers_1_attention_self_key_weight_to_fp16_palettized, x = input_31_cast_fp16)[name = string("linear_7_cast_fp16")];
+            tensor<int32, [4]> var_350 = const()[name = string("op_350"), val = tensor<int32, [4]>([1, 256, 12, 32])];
+            tensor<fp16, [1, 256, 12, 32]> var_351_cast_fp16 = reshape(shape = var_350, x = linear_7_cast_fp16)[name = string("op_351_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_1_attention_self_value_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25236352))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25347008))))[name = string("layers_1_attention_self_value_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_1_attention_self_value_bias_to_fp16 = const()[name = string("layers_1_attention_self_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25350144)))];
+            tensor<fp16, [1, 256, 384]> linear_8_cast_fp16 = linear(bias = layers_1_attention_self_value_bias_to_fp16, weight = layers_1_attention_self_value_weight_to_fp16_palettized, x = input_31_cast_fp16)[name = string("linear_8_cast_fp16")];
+            tensor<int32, [4]> var_365 = const()[name = string("op_365"), val = tensor<int32, [4]>([1, 256, 12, 32])];
+            tensor<fp16, [1, 256, 12, 32]> var_366_cast_fp16 = reshape(shape = var_365, x = linear_8_cast_fp16)[name = string("op_366_cast_fp16")];
+            tensor<int32, [4]> var_371 = const()[name = string("op_371"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool var_376_transpose_x_0 = const()[name = string("op_376_transpose_x_0"), val = bool(false)];
+            bool var_376_transpose_y_0 = const()[name = string("op_376_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_26_perm_0 = const()[name = string("transpose_26_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_27_perm_0 = const()[name = string("transpose_27_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 12, 32, 256]> transpose_27 = transpose(perm = transpose_27_perm_0, x = var_351_cast_fp16)[name = string("transpose_54")];
+            tensor<fp16, [1, 12, 256, 32]> transpose_26 = transpose(perm = transpose_26_perm_0, x = var_336_cast_fp16)[name = string("transpose_55")];
+            tensor<fp16, [1, 12, 256, 256]> var_376_cast_fp16 = matmul(transpose_x = var_376_transpose_x_0, transpose_y = var_376_transpose_y_0, x = transpose_26, y = transpose_27)[name = string("op_376_cast_fp16")];
+            fp16 var_377_to_fp16 = const()[name = string("op_377_to_fp16"), val = fp16(0x1.6ap-3)];
+            tensor<fp16, [1, 12, 256, 256]> weights_5_cast_fp16 = mul(x = var_376_cast_fp16, y = var_377_to_fp16)[name = string("weights_5_cast_fp16")];
+            tensor<fp16, [1, 12, 256, 256]> var_380_cast_fp16 = add(x = weights_5_cast_fp16, y = additive_mask_cast_fp16)[name = string("op_380_cast_fp16")];
+            int32 var_381 = const()[name = string("op_381"), val = int32(-1)];
+            tensor<fp16, [1, 12, 256, 256]> input_33_cast_fp16 = softmax(axis = var_381, x = var_380_cast_fp16)[name = string("input_33_cast_fp16")];
+            bool var_387_transpose_x_0 = const()[name = string("op_387_transpose_x_0"), val = bool(false)];
+            bool var_387_transpose_y_0 = const()[name = string("op_387_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 256, 32]> value_3_cast_fp16 = transpose(perm = var_371, x = var_366_cast_fp16)[name = string("transpose_53")];
+            tensor<fp16, [1, 12, 256, 32]> var_387_cast_fp16 = matmul(transpose_x = var_387_transpose_x_0, transpose_y = var_387_transpose_y_0, x = input_33_cast_fp16, y = value_3_cast_fp16)[name = string("op_387_cast_fp16")];
+            tensor<int32, [4]> var_392 = const()[name = string("op_392"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_397 = const()[name = string("op_397"), val = tensor<int32, [3]>([1, 256, 384])];
+            tensor<fp16, [1, 256, 12, 32]> var_393_cast_fp16 = transpose(perm = var_392, x = var_387_cast_fp16)[name = string("transpose_52")];
+            tensor<fp16, [1, 256, 384]> input_35_cast_fp16 = reshape(shape = var_397, x = var_393_cast_fp16)[name = string("input_35_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_1_attention_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25350976))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25461632))))[name = string("layers_1_attention_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_1_attention_output_dense_bias_to_fp16 = const()[name = string("layers_1_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25464768)))];
+            tensor<fp16, [1, 256, 384]> linear_9_cast_fp16 = linear(bias = layers_1_attention_output_dense_bias_to_fp16, weight = layers_1_attention_output_dense_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = string("linear_9_cast_fp16")];
+            tensor<fp16, [1, 256, 384]> input_39_cast_fp16 = add(x = linear_9_cast_fp16, y = input_31_cast_fp16)[name = string("input_39_cast_fp16")];
+            tensor<int32, [1]> input_41_axes_0 = const()[name = string("input_41_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_1_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_1_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25465600)))];
+            tensor<fp16, [384]> layers_1_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_1_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25466432)))];
+            fp16 var_400_to_fp16 = const()[name = string("op_400_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 256, 384]> input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = layers_1_attention_output_LayerNorm_bias_to_fp16, epsilon = var_400_to_fp16, gamma = layers_1_attention_output_LayerNorm_weight_to_fp16, x = input_39_cast_fp16)[name = string("input_41_cast_fp16")];
+            tensor<fp16, [1536, 384]> layers_1_intermediate_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25467264))), lut = tensor<fp16, [96, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25909696))))[name = string("layers_1_intermediate_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [1536]> layers_1_intermediate_dense_bias_to_fp16 = const()[name = string("layers_1_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25922048)))];
+            tensor<fp16, [1, 256, 1536]> linear_10_cast_fp16 = linear(bias = layers_1_intermediate_dense_bias_to_fp16, weight = layers_1_intermediate_dense_weight_to_fp16_palettized, x = input_41_cast_fp16)[name = string("linear_10_cast_fp16")];
+            string input_45_mode_0 = const()[name = string("input_45_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 256, 1536]> input_45_cast_fp16 = gelu(mode = input_45_mode_0, x = linear_10_cast_fp16)[name = string("input_45_cast_fp16")];
+            tensor<fp16, [384, 1536]> layers_1_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25925184))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26367616))))[name = string("layers_1_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_1_output_dense_bias_to_fp16 = const()[name = string("layers_1_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26370752)))];
+            tensor<fp16, [1, 256, 384]> linear_11_cast_fp16 = linear(bias = layers_1_output_dense_bias_to_fp16, weight = layers_1_output_dense_weight_to_fp16_palettized, x = input_45_cast_fp16)[name = string("linear_11_cast_fp16")];
+            tensor<fp16, [1, 256, 384]> input_49_cast_fp16 = add(x = linear_11_cast_fp16, y = input_41_cast_fp16)[name = string("input_49_cast_fp16")];
+            tensor<int32, [1]> input_51_axes_0 = const()[name = string("input_51_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_1_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_1_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26371584)))];
+            tensor<fp16, [384]> layers_1_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_1_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26372416)))];
+            fp16 var_423_to_fp16 = const()[name = string("op_423_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 256, 384]> input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = layers_1_output_LayerNorm_bias_to_fp16, epsilon = var_423_to_fp16, gamma = layers_1_output_LayerNorm_weight_to_fp16, x = input_49_cast_fp16)[name = string("input_51_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_2_attention_self_query_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26373248))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26483904))))[name = string("layers_2_attention_self_query_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_2_attention_self_query_bias_to_fp16 = const()[name = string("layers_2_attention_self_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26487040)))];
+            tensor<fp16, [1, 256, 384]> linear_12_cast_fp16 = linear(bias = layers_2_attention_self_query_bias_to_fp16, weight = layers_2_attention_self_query_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = string("linear_12_cast_fp16")];
+            tensor<int32, [4]> var_446 = const()[name = string("op_446"), val = tensor<int32, [4]>([1, 256, 12, 32])];
+            tensor<fp16, [1, 256, 12, 32]> var_447_cast_fp16 = reshape(shape = var_446, x = linear_12_cast_fp16)[name = string("op_447_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_2_attention_self_key_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26487872))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26598528))))[name = string("layers_2_attention_self_key_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_2_attention_self_key_bias_to_fp16 = const()[name = string("layers_2_attention_self_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26601664)))];
+            tensor<fp16, [1, 256, 384]> linear_13_cast_fp16 = linear(bias = layers_2_attention_self_key_bias_to_fp16, weight = layers_2_attention_self_key_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = string("linear_13_cast_fp16")];
+            tensor<int32, [4]> var_461 = const()[name = string("op_461"), val = tensor<int32, [4]>([1, 256, 12, 32])];
+            tensor<fp16, [1, 256, 12, 32]> var_462_cast_fp16 = reshape(shape = var_461, x = linear_13_cast_fp16)[name = string("op_462_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_2_attention_self_value_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26602496))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26713152))))[name = string("layers_2_attention_self_value_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_2_attention_self_value_bias_to_fp16 = const()[name = string("layers_2_attention_self_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26716288)))];
+            tensor<fp16, [1, 256, 384]> linear_14_cast_fp16 = linear(bias = layers_2_attention_self_value_bias_to_fp16, weight = layers_2_attention_self_value_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = string("linear_14_cast_fp16")];
+            tensor<int32, [4]> var_476 = const()[name = string("op_476"), val = tensor<int32, [4]>([1, 256, 12, 32])];
+            tensor<fp16, [1, 256, 12, 32]> var_477_cast_fp16 = reshape(shape = var_476, x = linear_14_cast_fp16)[name = string("op_477_cast_fp16")];
+            tensor<int32, [4]> var_482 = const()[name = string("op_482"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool var_487_transpose_x_0 = const()[name = string("op_487_transpose_x_0"), val = bool(false)];
+            bool var_487_transpose_y_0 = const()[name = string("op_487_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_28_perm_0 = const()[name = string("transpose_28_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_29_perm_0 = const()[name = string("transpose_29_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 12, 32, 256]> transpose_29 = transpose(perm = transpose_29_perm_0, x = var_462_cast_fp16)[name = string("transpose_50")];
+            tensor<fp16, [1, 12, 256, 32]> transpose_28 = transpose(perm = transpose_28_perm_0, x = var_447_cast_fp16)[name = string("transpose_51")];
+            tensor<fp16, [1, 12, 256, 256]> var_487_cast_fp16 = matmul(transpose_x = var_487_transpose_x_0, transpose_y = var_487_transpose_y_0, x = transpose_28, y = transpose_29)[name = string("op_487_cast_fp16")];
+            fp16 var_488_to_fp16 = const()[name = string("op_488_to_fp16"), val = fp16(0x1.6ap-3)];
+            tensor<fp16, [1, 12, 256, 256]> weights_9_cast_fp16 = mul(x = var_487_cast_fp16, y = var_488_to_fp16)[name = string("weights_9_cast_fp16")];
+            tensor<fp16, [1, 12, 256, 256]> var_491_cast_fp16 = add(x = weights_9_cast_fp16, y = additive_mask_cast_fp16)[name = string("op_491_cast_fp16")];
+            int32 var_492 = const()[name = string("op_492"), val = int32(-1)];
+            tensor<fp16, [1, 12, 256, 256]> input_53_cast_fp16 = softmax(axis = var_492, x = var_491_cast_fp16)[name = string("input_53_cast_fp16")];
+            bool var_498_transpose_x_0 = const()[name = string("op_498_transpose_x_0"), val = bool(false)];
+            bool var_498_transpose_y_0 = const()[name = string("op_498_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 256, 32]> value_5_cast_fp16 = transpose(perm = var_482, x = var_477_cast_fp16)[name = string("transpose_49")];
+            tensor<fp16, [1, 12, 256, 32]> var_498_cast_fp16 = matmul(transpose_x = var_498_transpose_x_0, transpose_y = var_498_transpose_y_0, x = input_53_cast_fp16, y = value_5_cast_fp16)[name = string("op_498_cast_fp16")];
+            tensor<int32, [4]> var_503 = const()[name = string("op_503"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_508 = const()[name = string("op_508"), val = tensor<int32, [3]>([1, 256, 384])];
+            tensor<fp16, [1, 256, 12, 32]> var_504_cast_fp16 = transpose(perm = var_503, x = var_498_cast_fp16)[name = string("transpose_48")];
+            tensor<fp16, [1, 256, 384]> input_55_cast_fp16 = reshape(shape = var_508, x = var_504_cast_fp16)[name = string("input_55_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_2_attention_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26717120))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26827776))))[name = string("layers_2_attention_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_2_attention_output_dense_bias_to_fp16 = const()[name = string("layers_2_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26830912)))];
+            tensor<fp16, [1, 256, 384]> linear_15_cast_fp16 = linear(bias = layers_2_attention_output_dense_bias_to_fp16, weight = layers_2_attention_output_dense_weight_to_fp16_palettized, x = input_55_cast_fp16)[name = string("linear_15_cast_fp16")];
+            tensor<fp16, [1, 256, 384]> input_59_cast_fp16 = add(x = linear_15_cast_fp16, y = input_51_cast_fp16)[name = string("input_59_cast_fp16")];
+            tensor<int32, [1]> input_61_axes_0 = const()[name = string("input_61_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_2_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_2_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26831744)))];
+            tensor<fp16, [384]> layers_2_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_2_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26832576)))];
+            fp16 var_511_to_fp16 = const()[name = string("op_511_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 256, 384]> input_61_cast_fp16 = layer_norm(axes = input_61_axes_0, beta = layers_2_attention_output_LayerNorm_bias_to_fp16, epsilon = var_511_to_fp16, gamma = layers_2_attention_output_LayerNorm_weight_to_fp16, x = input_59_cast_fp16)[name = string("input_61_cast_fp16")];
+            tensor<fp16, [1536, 384]> layers_2_intermediate_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26833408))), lut = tensor<fp16, [96, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27275840))))[name = string("layers_2_intermediate_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [1536]> layers_2_intermediate_dense_bias_to_fp16 = const()[name = string("layers_2_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27288192)))];
+            tensor<fp16, [1, 256, 1536]> linear_16_cast_fp16 = linear(bias = layers_2_intermediate_dense_bias_to_fp16, weight = layers_2_intermediate_dense_weight_to_fp16_palettized, x = input_61_cast_fp16)[name = string("linear_16_cast_fp16")];
+            string input_65_mode_0 = const()[name = string("input_65_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 256, 1536]> input_65_cast_fp16 = gelu(mode = input_65_mode_0, x = linear_16_cast_fp16)[name = string("input_65_cast_fp16")];
+            tensor<fp16, [384, 1536]> layers_2_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27291328))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27733760))))[name = string("layers_2_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_2_output_dense_bias_to_fp16 = const()[name = string("layers_2_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27736896)))];
+            tensor<fp16, [1, 256, 384]> linear_17_cast_fp16 = linear(bias = layers_2_output_dense_bias_to_fp16, weight = layers_2_output_dense_weight_to_fp16_palettized, x = input_65_cast_fp16)[name = string("linear_17_cast_fp16")];
+            tensor<fp16, [1, 256, 384]> input_69_cast_fp16 = add(x = linear_17_cast_fp16, y = input_61_cast_fp16)[name = string("input_69_cast_fp16")];
+            tensor<int32, [1]> input_71_axes_0 = const()[name = string("input_71_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_2_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_2_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27737728)))];
+            tensor<fp16, [384]> layers_2_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_2_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27738560)))];
+            fp16 var_534_to_fp16 = const()[name = string("op_534_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 256, 384]> input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = layers_2_output_LayerNorm_bias_to_fp16, epsilon = var_534_to_fp16, gamma = layers_2_output_LayerNorm_weight_to_fp16, x = input_69_cast_fp16)[name = string("input_71_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_3_attention_self_query_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27739392))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27850048))))[name = string("layers_3_attention_self_query_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_3_attention_self_query_bias_to_fp16 = const()[name = string("layers_3_attention_self_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27853184)))];
+            tensor<fp16, [1, 256, 384]> linear_18_cast_fp16 = linear(bias = layers_3_attention_self_query_bias_to_fp16, weight = layers_3_attention_self_query_weight_to_fp16_palettized, x = input_71_cast_fp16)[name = string("linear_18_cast_fp16")];
+            tensor<int32, [4]> var_557 = const()[name = string("op_557"), val = tensor<int32, [4]>([1, 256, 12, 32])];
+            tensor<fp16, [1, 256, 12, 32]> var_558_cast_fp16 = reshape(shape = var_557, x = linear_18_cast_fp16)[name = string("op_558_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_3_attention_self_key_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27854016))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27964672))))[name = string("layers_3_attention_self_key_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_3_attention_self_key_bias_to_fp16 = const()[name = string("layers_3_attention_self_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27967808)))];
+            tensor<fp16, [1, 256, 384]> linear_19_cast_fp16 = linear(bias = layers_3_attention_self_key_bias_to_fp16, weight = layers_3_attention_self_key_weight_to_fp16_palettized, x = input_71_cast_fp16)[name = string("linear_19_cast_fp16")];
+            tensor<int32, [4]> var_572 = const()[name = string("op_572"), val = tensor<int32, [4]>([1, 256, 12, 32])];
+            tensor<fp16, [1, 256, 12, 32]> var_573_cast_fp16 = reshape(shape = var_572, x = linear_19_cast_fp16)[name = string("op_573_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_3_attention_self_value_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27968640))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28079296))))[name = string("layers_3_attention_self_value_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_3_attention_self_value_bias_to_fp16 = const()[name = string("layers_3_attention_self_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28082432)))];
+            tensor<fp16, [1, 256, 384]> linear_20_cast_fp16 = linear(bias = layers_3_attention_self_value_bias_to_fp16, weight = layers_3_attention_self_value_weight_to_fp16_palettized, x = input_71_cast_fp16)[name = string("linear_20_cast_fp16")];
+            tensor<int32, [4]> var_587 = const()[name = string("op_587"), val = tensor<int32, [4]>([1, 256, 12, 32])];
+            tensor<fp16, [1, 256, 12, 32]> var_588_cast_fp16 = reshape(shape = var_587, x = linear_20_cast_fp16)[name = string("op_588_cast_fp16")];
+            tensor<int32, [4]> var_593 = const()[name = string("op_593"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool var_598_transpose_x_0 = const()[name = string("op_598_transpose_x_0"), val = bool(false)];
+            bool var_598_transpose_y_0 = const()[name = string("op_598_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_30_perm_0 = const()[name = string("transpose_30_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_31_perm_0 = const()[name = string("transpose_31_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 12, 32, 256]> transpose_31 = transpose(perm = transpose_31_perm_0, x = var_573_cast_fp16)[name = string("transpose_46")];
+            tensor<fp16, [1, 12, 256, 32]> transpose_30 = transpose(perm = transpose_30_perm_0, x = var_558_cast_fp16)[name = string("transpose_47")];
+            tensor<fp16, [1, 12, 256, 256]> var_598_cast_fp16 = matmul(transpose_x = var_598_transpose_x_0, transpose_y = var_598_transpose_y_0, x = transpose_30, y = transpose_31)[name = string("op_598_cast_fp16")];
+            fp16 var_599_to_fp16 = const()[name = string("op_599_to_fp16"), val = fp16(0x1.6ap-3)];
+            tensor<fp16, [1, 12, 256, 256]> weights_13_cast_fp16 = mul(x = var_598_cast_fp16, y = var_599_to_fp16)[name = string("weights_13_cast_fp16")];
+            tensor<fp16, [1, 12, 256, 256]> var_602_cast_fp16 = add(x = weights_13_cast_fp16, y = additive_mask_cast_fp16)[name = string("op_602_cast_fp16")];
+            int32 var_603 = const()[name = string("op_603"), val = int32(-1)];
+            tensor<fp16, [1, 12, 256, 256]> input_73_cast_fp16 = softmax(axis = var_603, x = var_602_cast_fp16)[name = string("input_73_cast_fp16")];
+            bool var_609_transpose_x_0 = const()[name = string("op_609_transpose_x_0"), val = bool(false)];
+            bool var_609_transpose_y_0 = const()[name = string("op_609_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 256, 32]> value_7_cast_fp16 = transpose(perm = var_593, x = var_588_cast_fp16)[name = string("transpose_45")];
+            tensor<fp16, [1, 12, 256, 32]> var_609_cast_fp16 = matmul(transpose_x = var_609_transpose_x_0, transpose_y = var_609_transpose_y_0, x = input_73_cast_fp16, y = value_7_cast_fp16)[name = string("op_609_cast_fp16")];
+            tensor<int32, [4]> var_614 = const()[name = string("op_614"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_619 = const()[name = string("op_619"), val = tensor<int32, [3]>([1, 256, 384])];
+            tensor<fp16, [1, 256, 12, 32]> var_615_cast_fp16 = transpose(perm = var_614, x = var_609_cast_fp16)[name = string("transpose_44")];
+            tensor<fp16, [1, 256, 384]> input_75_cast_fp16 = reshape(shape = var_619, x = var_615_cast_fp16)[name = string("input_75_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_3_attention_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28083264))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28193920))))[name = string("layers_3_attention_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_3_attention_output_dense_bias_to_fp16 = const()[name = string("layers_3_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28197056)))];
+            tensor<fp16, [1, 256, 384]> linear_21_cast_fp16 = linear(bias = layers_3_attention_output_dense_bias_to_fp16, weight = layers_3_attention_output_dense_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = string("linear_21_cast_fp16")];
+            tensor<fp16, [1, 256, 384]> input_79_cast_fp16 = add(x = linear_21_cast_fp16, y = input_71_cast_fp16)[name = string("input_79_cast_fp16")];
+            tensor<int32, [1]> input_81_axes_0 = const()[name = string("input_81_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_3_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_3_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28197888)))];
+            tensor<fp16, [384]> layers_3_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_3_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28198720)))];
+            fp16 var_622_to_fp16 = const()[name = string("op_622_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 256, 384]> input_81_cast_fp16 = layer_norm(axes = input_81_axes_0, beta = layers_3_attention_output_LayerNorm_bias_to_fp16, epsilon = var_622_to_fp16, gamma = layers_3_attention_output_LayerNorm_weight_to_fp16, x = input_79_cast_fp16)[name = string("input_81_cast_fp16")];
+            tensor<fp16, [1536, 384]> layers_3_intermediate_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28199552))), lut = tensor<fp16, [96, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28641984))))[name = string("layers_3_intermediate_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [1536]> layers_3_intermediate_dense_bias_to_fp16 = const()[name = string("layers_3_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28654336)))];
+            tensor<fp16, [1, 256, 1536]> linear_22_cast_fp16 = linear(bias = layers_3_intermediate_dense_bias_to_fp16, weight = layers_3_intermediate_dense_weight_to_fp16_palettized, x = input_81_cast_fp16)[name = string("linear_22_cast_fp16")];
+            string input_85_mode_0 = const()[name = string("input_85_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 256, 1536]> input_85_cast_fp16 = gelu(mode = input_85_mode_0, x = linear_22_cast_fp16)[name = string("input_85_cast_fp16")];
+            tensor<fp16, [384, 1536]> layers_3_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28657472))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29099904))))[name = string("layers_3_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_3_output_dense_bias_to_fp16 = const()[name = string("layers_3_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29103040)))];
+            tensor<fp16, [1, 256, 384]> linear_23_cast_fp16 = linear(bias = layers_3_output_dense_bias_to_fp16, weight = layers_3_output_dense_weight_to_fp16_palettized, x = input_85_cast_fp16)[name = string("linear_23_cast_fp16")];
+            tensor<fp16, [1, 256, 384]> input_89_cast_fp16 = add(x = linear_23_cast_fp16, y = input_81_cast_fp16)[name = string("input_89_cast_fp16")];
+            tensor<int32, [1]> input_91_axes_0 = const()[name = string("input_91_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_3_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_3_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29103872)))];
+            tensor<fp16, [384]> layers_3_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_3_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29104704)))];
+            fp16 var_645_to_fp16 = const()[name = string("op_645_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 256, 384]> input_91_cast_fp16 = layer_norm(axes = input_91_axes_0, beta = layers_3_output_LayerNorm_bias_to_fp16, epsilon = var_645_to_fp16, gamma = layers_3_output_LayerNorm_weight_to_fp16, x = input_89_cast_fp16)[name = string("input_91_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_4_attention_self_query_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29105536))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29216192))))[name = string("layers_4_attention_self_query_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_4_attention_self_query_bias_to_fp16 = const()[name = string("layers_4_attention_self_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29219328)))];
+            tensor<fp16, [1, 256, 384]> linear_24_cast_fp16 = linear(bias = layers_4_attention_self_query_bias_to_fp16, weight = layers_4_attention_self_query_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = string("linear_24_cast_fp16")];
+            tensor<int32, [4]> var_668 = const()[name = string("op_668"), val = tensor<int32, [4]>([1, 256, 12, 32])];
+            tensor<fp16, [1, 256, 12, 32]> var_669_cast_fp16 = reshape(shape = var_668, x = linear_24_cast_fp16)[name = string("op_669_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_4_attention_self_key_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29220160))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29330816))))[name = string("layers_4_attention_self_key_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_4_attention_self_key_bias_to_fp16 = const()[name = string("layers_4_attention_self_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29333952)))];
+            tensor<fp16, [1, 256, 384]> linear_25_cast_fp16 = linear(bias = layers_4_attention_self_key_bias_to_fp16, weight = layers_4_attention_self_key_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = string("linear_25_cast_fp16")];
+            tensor<int32, [4]> var_683 = const()[name = string("op_683"), val = tensor<int32, [4]>([1, 256, 12, 32])];
+            tensor<fp16, [1, 256, 12, 32]> var_684_cast_fp16 = reshape(shape = var_683, x = linear_25_cast_fp16)[name = string("op_684_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_4_attention_self_value_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29334784))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29445440))))[name = string("layers_4_attention_self_value_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_4_attention_self_value_bias_to_fp16 = const()[name = string("layers_4_attention_self_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29448576)))];
+            tensor<fp16, [1, 256, 384]> linear_26_cast_fp16 = linear(bias = layers_4_attention_self_value_bias_to_fp16, weight = layers_4_attention_self_value_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = string("linear_26_cast_fp16")];
+            tensor<int32, [4]> var_698 = const()[name = string("op_698"), val = tensor<int32, [4]>([1, 256, 12, 32])];
+            tensor<fp16, [1, 256, 12, 32]> var_699_cast_fp16 = reshape(shape = var_698, x = linear_26_cast_fp16)[name = string("op_699_cast_fp16")];
+            tensor<int32, [4]> var_704 = const()[name = string("op_704"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool var_709_transpose_x_0 = const()[name = string("op_709_transpose_x_0"), val = bool(false)];
+            bool var_709_transpose_y_0 = const()[name = string("op_709_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_32_perm_0 = const()[name = string("transpose_32_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_33_perm_0 = const()[name = string("transpose_33_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 12, 32, 256]> transpose_33 = transpose(perm = transpose_33_perm_0, x = var_684_cast_fp16)[name = string("transpose_42")];
+            tensor<fp16, [1, 12, 256, 32]> transpose_32 = transpose(perm = transpose_32_perm_0, x = var_669_cast_fp16)[name = string("transpose_43")];
+            tensor<fp16, [1, 12, 256, 256]> var_709_cast_fp16 = matmul(transpose_x = var_709_transpose_x_0, transpose_y = var_709_transpose_y_0, x = transpose_32, y = transpose_33)[name = string("op_709_cast_fp16")];
+            fp16 var_710_to_fp16 = const()[name = string("op_710_to_fp16"), val = fp16(0x1.6ap-3)];
+            tensor<fp16, [1, 12, 256, 256]> weights_17_cast_fp16 = mul(x = var_709_cast_fp16, y = var_710_to_fp16)[name = string("weights_17_cast_fp16")];
+            tensor<fp16, [1, 12, 256, 256]> var_713_cast_fp16 = add(x = weights_17_cast_fp16, y = additive_mask_cast_fp16)[name = string("op_713_cast_fp16")];
+            int32 var_714 = const()[name = string("op_714"), val = int32(-1)];
+            tensor<fp16, [1, 12, 256, 256]> input_93_cast_fp16 = softmax(axis = var_714, x = var_713_cast_fp16)[name = string("input_93_cast_fp16")];
+            bool var_720_transpose_x_0 = const()[name = string("op_720_transpose_x_0"), val = bool(false)];
+            bool var_720_transpose_y_0 = const()[name = string("op_720_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 256, 32]> value_9_cast_fp16 = transpose(perm = var_704, x = var_699_cast_fp16)[name = string("transpose_41")];
+            tensor<fp16, [1, 12, 256, 32]> var_720_cast_fp16 = matmul(transpose_x = var_720_transpose_x_0, transpose_y = var_720_transpose_y_0, x = input_93_cast_fp16, y = value_9_cast_fp16)[name = string("op_720_cast_fp16")];
+            tensor<int32, [4]> var_725 = const()[name = string("op_725"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_730 = const()[name = string("op_730"), val = tensor<int32, [3]>([1, 256, 384])];
+            tensor<fp16, [1, 256, 12, 32]> var_726_cast_fp16 = transpose(perm = var_725, x = var_720_cast_fp16)[name = string("transpose_40")];
+            tensor<fp16, [1, 256, 384]> input_95_cast_fp16 = reshape(shape = var_730, x = var_726_cast_fp16)[name = string("input_95_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_4_attention_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29449408))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29560064))))[name = string("layers_4_attention_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_4_attention_output_dense_bias_to_fp16 = const()[name = string("layers_4_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29563200)))];
+            tensor<fp16, [1, 256, 384]> linear_27_cast_fp16 = linear(bias = layers_4_attention_output_dense_bias_to_fp16, weight = layers_4_attention_output_dense_weight_to_fp16_palettized, x = input_95_cast_fp16)[name = string("linear_27_cast_fp16")];
+            tensor<fp16, [1, 256, 384]> input_99_cast_fp16 = add(x = linear_27_cast_fp16, y = input_91_cast_fp16)[name = string("input_99_cast_fp16")];
+            tensor<int32, [1]> input_101_axes_0 = const()[name = string("input_101_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_4_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_4_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29564032)))];
+            tensor<fp16, [384]> layers_4_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_4_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29564864)))];
+            fp16 var_733_to_fp16 = const()[name = string("op_733_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 256, 384]> input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = layers_4_attention_output_LayerNorm_bias_to_fp16, epsilon = var_733_to_fp16, gamma = layers_4_attention_output_LayerNorm_weight_to_fp16, x = input_99_cast_fp16)[name = string("input_101_cast_fp16")];
+            tensor<fp16, [1536, 384]> layers_4_intermediate_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29565696))), lut = tensor<fp16, [96, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30008128))))[name = string("layers_4_intermediate_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [1536]> layers_4_intermediate_dense_bias_to_fp16 = const()[name = string("layers_4_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30020480)))];
+            tensor<fp16, [1, 256, 1536]> linear_28_cast_fp16 = linear(bias = layers_4_intermediate_dense_bias_to_fp16, weight = layers_4_intermediate_dense_weight_to_fp16_palettized, x = input_101_cast_fp16)[name = string("linear_28_cast_fp16")];
+            string input_105_mode_0 = const()[name = string("input_105_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 256, 1536]> input_105_cast_fp16 = gelu(mode = input_105_mode_0, x = linear_28_cast_fp16)[name = string("input_105_cast_fp16")];
+            tensor<fp16, [384, 1536]> layers_4_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30023616))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30466048))))[name = string("layers_4_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_4_output_dense_bias_to_fp16 = const()[name = string("layers_4_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30469184)))];
+            tensor<fp16, [1, 256, 384]> linear_29_cast_fp16 = linear(bias = layers_4_output_dense_bias_to_fp16, weight = layers_4_output_dense_weight_to_fp16_palettized, x = input_105_cast_fp16)[name = string("linear_29_cast_fp16")];
+            tensor<fp16, [1, 256, 384]> input_109_cast_fp16 = add(x = linear_29_cast_fp16, y = input_101_cast_fp16)[name = string("input_109_cast_fp16")];
+            tensor<int32, [1]> input_111_axes_0 = const()[name = string("input_111_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_4_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_4_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30470016)))];
+            tensor<fp16, [384]> layers_4_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_4_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30470848)))];
+            fp16 var_756_to_fp16 = const()[name = string("op_756_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 256, 384]> input_111_cast_fp16 = layer_norm(axes = input_111_axes_0, beta = layers_4_output_LayerNorm_bias_to_fp16, epsilon = var_756_to_fp16, gamma = layers_4_output_LayerNorm_weight_to_fp16, x = input_109_cast_fp16)[name = string("input_111_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_5_attention_self_query_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30471680))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30582336))))[name = string("layers_5_attention_self_query_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_5_attention_self_query_bias_to_fp16 = const()[name = string("layers_5_attention_self_query_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30585472)))];
+            tensor<fp16, [1, 256, 384]> linear_30_cast_fp16 = linear(bias = layers_5_attention_self_query_bias_to_fp16, weight = layers_5_attention_self_query_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = string("linear_30_cast_fp16")];
+            tensor<int32, [4]> var_779 = const()[name = string("op_779"), val = tensor<int32, [4]>([1, 256, 12, 32])];
+            tensor<fp16, [1, 256, 12, 32]> var_780_cast_fp16 = reshape(shape = var_779, x = linear_30_cast_fp16)[name = string("op_780_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_5_attention_self_key_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30586304))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30696960))))[name = string("layers_5_attention_self_key_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_5_attention_self_key_bias_to_fp16 = const()[name = string("layers_5_attention_self_key_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30700096)))];
+            tensor<fp16, [1, 256, 384]> linear_31_cast_fp16 = linear(bias = layers_5_attention_self_key_bias_to_fp16, weight = layers_5_attention_self_key_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = string("linear_31_cast_fp16")];
+            tensor<int32, [4]> var_794 = const()[name = string("op_794"), val = tensor<int32, [4]>([1, 256, 12, 32])];
+            tensor<fp16, [1, 256, 12, 32]> var_795_cast_fp16 = reshape(shape = var_794, x = linear_31_cast_fp16)[name = string("op_795_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_5_attention_self_value_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30700928))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30811584))))[name = string("layers_5_attention_self_value_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_5_attention_self_value_bias_to_fp16 = const()[name = string("layers_5_attention_self_value_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30814720)))];
+            tensor<fp16, [1, 256, 384]> linear_32_cast_fp16 = linear(bias = layers_5_attention_self_value_bias_to_fp16, weight = layers_5_attention_self_value_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = string("linear_32_cast_fp16")];
+            tensor<int32, [4]> var_809 = const()[name = string("op_809"), val = tensor<int32, [4]>([1, 256, 12, 32])];
+            tensor<fp16, [1, 256, 12, 32]> var_810_cast_fp16 = reshape(shape = var_809, x = linear_32_cast_fp16)[name = string("op_810_cast_fp16")];
+            tensor<int32, [4]> var_815 = const()[name = string("op_815"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            bool var_820_transpose_x_0 = const()[name = string("op_820_transpose_x_0"), val = bool(false)];
+            bool var_820_transpose_y_0 = const()[name = string("op_820_transpose_y_0"), val = bool(false)];
+            tensor<int32, [4]> transpose_34_perm_0 = const()[name = string("transpose_34_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_35_perm_0 = const()[name = string("transpose_35_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp16, [1, 12, 32, 256]> transpose_35 = transpose(perm = transpose_35_perm_0, x = var_795_cast_fp16)[name = string("transpose_38")];
+            tensor<fp16, [1, 12, 256, 32]> transpose_34 = transpose(perm = transpose_34_perm_0, x = var_780_cast_fp16)[name = string("transpose_39")];
+            tensor<fp16, [1, 12, 256, 256]> var_820_cast_fp16 = matmul(transpose_x = var_820_transpose_x_0, transpose_y = var_820_transpose_y_0, x = transpose_34, y = transpose_35)[name = string("op_820_cast_fp16")];
+            fp16 var_821_to_fp16 = const()[name = string("op_821_to_fp16"), val = fp16(0x1.6ap-3)];
+            tensor<fp16, [1, 12, 256, 256]> weights_21_cast_fp16 = mul(x = var_820_cast_fp16, y = var_821_to_fp16)[name = string("weights_21_cast_fp16")];
+            tensor<fp16, [1, 12, 256, 256]> var_824_cast_fp16 = add(x = weights_21_cast_fp16, y = additive_mask_cast_fp16)[name = string("op_824_cast_fp16")];
+            int32 var_825 = const()[name = string("op_825"), val = int32(-1)];
+            tensor<fp16, [1, 12, 256, 256]> input_113_cast_fp16 = softmax(axis = var_825, x = var_824_cast_fp16)[name = string("input_113_cast_fp16")];
+            bool var_831_transpose_x_0 = const()[name = string("op_831_transpose_x_0"), val = bool(false)];
+            bool var_831_transpose_y_0 = const()[name = string("op_831_transpose_y_0"), val = bool(false)];
+            tensor<fp16, [1, 12, 256, 32]> value_11_cast_fp16 = transpose(perm = var_815, x = var_810_cast_fp16)[name = string("transpose_37")];
+            tensor<fp16, [1, 12, 256, 32]> var_831_cast_fp16 = matmul(transpose_x = var_831_transpose_x_0, transpose_y = var_831_transpose_y_0, x = input_113_cast_fp16, y = value_11_cast_fp16)[name = string("op_831_cast_fp16")];
+            tensor<int32, [4]> var_836 = const()[name = string("op_836"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_841 = const()[name = string("op_841"), val = tensor<int32, [3]>([1, 256, 384])];
+            tensor<fp16, [1, 256, 12, 32]> var_837_cast_fp16 = transpose(perm = var_836, x = var_831_cast_fp16)[name = string("transpose_36")];
+            tensor<fp16, [1, 256, 384]> input_115_cast_fp16 = reshape(shape = var_841, x = var_837_cast_fp16)[name = string("input_115_cast_fp16")];
+            tensor<fp16, [384, 384]> layers_5_attention_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30815552))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30926208))))[name = string("layers_5_attention_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_5_attention_output_dense_bias_to_fp16 = const()[name = string("layers_5_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30929344)))];
+            tensor<fp16, [1, 256, 384]> linear_33_cast_fp16 = linear(bias = layers_5_attention_output_dense_bias_to_fp16, weight = layers_5_attention_output_dense_weight_to_fp16_palettized, x = input_115_cast_fp16)[name = string("linear_33_cast_fp16")];
+            tensor<fp16, [1, 256, 384]> input_119_cast_fp16 = add(x = linear_33_cast_fp16, y = input_111_cast_fp16)[name = string("input_119_cast_fp16")];
+            tensor<int32, [1]> input_121_axes_0 = const()[name = string("input_121_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_5_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_5_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30930176)))];
+            tensor<fp16, [384]> layers_5_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_5_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30931008)))];
+            fp16 var_844_to_fp16 = const()[name = string("op_844_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 256, 384]> input_121_cast_fp16 = layer_norm(axes = input_121_axes_0, beta = layers_5_attention_output_LayerNorm_bias_to_fp16, epsilon = var_844_to_fp16, gamma = layers_5_attention_output_LayerNorm_weight_to_fp16, x = input_119_cast_fp16)[name = string("input_121_cast_fp16")];
+            tensor<fp16, [1536, 384]> layers_5_intermediate_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30931840))), lut = tensor<fp16, [96, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31374272))))[name = string("layers_5_intermediate_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [1536]> layers_5_intermediate_dense_bias_to_fp16 = const()[name = string("layers_5_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31386624)))];
+            tensor<fp16, [1, 256, 1536]> linear_34_cast_fp16 = linear(bias = layers_5_intermediate_dense_bias_to_fp16, weight = layers_5_intermediate_dense_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = string("linear_34_cast_fp16")];
+            string input_125_mode_0 = const()[name = string("input_125_mode_0"), val = string("EXACT")];
+            tensor<fp16, [1, 256, 1536]> input_125_cast_fp16 = gelu(mode = input_125_mode_0, x = linear_34_cast_fp16)[name = string("input_125_cast_fp16")];
+            tensor<fp16, [384, 1536]> layers_5_output_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31389760))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31832192))))[name = string("layers_5_output_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> layers_5_output_dense_bias_to_fp16 = const()[name = string("layers_5_output_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31835328)))];
+            tensor<fp16, [1, 256, 384]> linear_35_cast_fp16 = linear(bias = layers_5_output_dense_bias_to_fp16, weight = layers_5_output_dense_weight_to_fp16_palettized, x = input_125_cast_fp16)[name = string("linear_35_cast_fp16")];
+            tensor<fp16, [1, 256, 384]> input_129_cast_fp16 = add(x = linear_35_cast_fp16, y = input_121_cast_fp16)[name = string("input_129_cast_fp16")];
+            tensor<int32, [1]> hidden_states_axes_0 = const()[name = string("hidden_states_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp16, [384]> layers_5_output_LayerNorm_weight_to_fp16 = const()[name = string("layers_5_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31836160)))];
+            tensor<fp16, [384]> layers_5_output_LayerNorm_bias_to_fp16 = const()[name = string("layers_5_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31836992)))];
+            fp16 var_867_to_fp16 = const()[name = string("op_867_to_fp16"), val = fp16(0x1p-24)];
+            tensor<fp16, [1, 256, 384]> hidden_states_cast_fp16 = layer_norm(axes = hidden_states_axes_0, beta = layers_5_output_LayerNorm_bias_to_fp16, epsilon = var_867_to_fp16, gamma = layers_5_output_LayerNorm_weight_to_fp16, x = input_129_cast_fp16)[name = string("hidden_states_cast_fp16")];
+            tensor<int32, [3]> input_131_begin_0 = const()[name = string("input_131_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
+            tensor<int32, [3]> input_131_end_0 = const()[name = string("input_131_end_0"), val = tensor<int32, [3]>([1, 1, 384])];
+            tensor<bool, [3]> input_131_end_mask_0 = const()[name = string("input_131_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
+            tensor<bool, [3]> input_131_squeeze_mask_0 = const()[name = string("input_131_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
+            tensor<fp16, [1, 384]> input_131_cast_fp16 = slice_by_index(begin = input_131_begin_0, end = input_131_end_0, end_mask = input_131_end_mask_0, squeeze_mask = input_131_squeeze_mask_0, x = hidden_states_cast_fp16)[name = string("input_131_cast_fp16")];
+            tensor<fp16, [384, 384]> pooler_dense_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31837824))), lut = tensor<fp16, [24, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31948480))))[name = string("pooler_dense_weight_to_fp16_palettized")];
+            tensor<fp16, [384]> pooler_dense_bias_to_fp16 = const()[name = string("pooler_dense_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31951616)))];
+            tensor<fp16, [1, 384]> linear_36_cast_fp16 = linear(bias = pooler_dense_bias_to_fp16, weight = pooler_dense_weight_to_fp16_palettized, x = input_131_cast_fp16)[name = string("linear_36_cast_fp16")];
+            tensor<fp16, [1, 384]> input_135_cast_fp16 = tanh(x = linear_36_cast_fp16)[name = string("input_135_cast_fp16")];
+            tensor<fp16, [1, 384]> classifier_weight_to_fp16 = const()[name = string("classifier_weight_to_fp16"), val = tensor<fp16, [1, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31952448)))];
+            tensor<fp16, [1]> classifier_bias_to_fp16 = const()[name = string("classifier_bias_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
+            tensor<fp16, [1, 1]> linear_37_cast_fp16 = linear(bias = classifier_bias_to_fp16, weight = classifier_weight_to_fp16, x = input_135_cast_fp16)[name = string("linear_37_cast_fp16")];
+            string linear_37_cast_fp16_to_fp32_dtype_0 = const()[name = string("linear_37_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
+            tensor<fp32, [1, 1]> score = cast(dtype = linear_37_cast_fp16_to_fp32_dtype_0, x = linear_37_cast_fp16)[name = string("cast_1")];
+        } -> (score);
+}

minilm_reranker_256.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8eb6432fef9c90fc527f8850ae1a3f86d76ab262b59ab040bb96ef94ce6964ca
+size 31953280