Upload 16 files

G2PDecoder.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:153360a0a799c3e48b56456324f057557b69dedf183ae3d137b45c95df5b15b2
-size 17997
+oid sha256:3b2bfc025be6563494c66b3f9c1b62674741e392d14d92d2ecaeb1730643e50c
+size 85095

G2PDecoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cbaeb4e743359f607ab161af0c6d8a817462fdaec622ee788ef8ef952c5f8214
-size 828030
+oid sha256:005b5ffbccabe4fd4074204d16af745b5d8f458a687b0c07e80060f7fe0a08e4
+size 25977088

G2PDecoder.mlpackage/Manifest.json

@@ -1,18 +1,18 @@
 {
     "fileFormatVersion": "1.0.0",
     "itemInfoEntries": {
-        "4AA5D221-1355-4F3F-B188-7F8EF536155C": {
-            "author": "com.apple.CoreML",
-            "description": "CoreML Model Specification",
-            "name": "model.mlmodel",
-            "path": "com.apple.CoreML/model.mlmodel"
-        },
-        "5D000863-3BC2-4F18-AC66-6F1DE5E29001": {
+        "837F74EA-4434-419F-AE52-FD61A69F6E0A": {
             "author": "com.apple.CoreML",
             "description": "CoreML Model Weights",
             "name": "weights",
             "path": "com.apple.CoreML/weights"
+        },
+        "FEF5E4E3-07B4-4F5A-BF3F-C4A2BB7E68E2": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
         }
     },
-    "rootModelIdentifier": "4AA5D221-1355-4F3F-B188-7F8EF536155C"
+    "rootModelIdentifier": "FEF5E4E3-07B4-4F5A-BF3F-C4A2BB7E68E2"
 }

G2PEncoder.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f51555b74730f0446406a50e3ee87dc4e2f0d759d2f095662880ac520a752fec
-size 20323
+oid sha256:338ee3222fb062efddd1f85374d1aa15478b4975b74f90c59264ad961d9879a3
+size 131789

G2PEncoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6926bcd2827d21fec82839487b987e06f85fd8a6a5bb896bc4f6062461d014ec
-size 694592
+oid sha256:11b225e93b6b1fe6fb85c9d4317724e1a033d6123fdc01dc2ad85f3c1f168eb3
+size 57056704

G2PEncoder.mlpackage/Manifest.json

@@ -1,18 +1,18 @@
 {
     "fileFormatVersion": "1.0.0",
     "itemInfoEntries": {
-        "C03F7353-6672-47CD-9842-1F68630789D7": {
-            "author": "com.apple.CoreML",
-            "description": "CoreML Model Weights",
-            "name": "weights",
-            "path": "com.apple.CoreML/weights"
-        },
-        "D40EF921-64C1-4EEC-BF85-D226D10F6B90": {
+        "48AE8C97-A0EB-4155-AE9E-D179548FBDD0": {
             "author": "com.apple.CoreML",
             "description": "CoreML Model Specification",
             "name": "model.mlmodel",
             "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "7760FC9D-D7D7-4E6E-8C85-EBF0D242EFC9": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
         }
     },
-    "rootModelIdentifier": "D40EF921-64C1-4EEC-BF85-D226D10F6B90"
+    "rootModelIdentifier": "48AE8C97-A0EB-4155-AE9E-D179548FBDD0"
 }

MultilingualG2PDecoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f34464b8c4e85cde3ca62bc7e22d1ab3eb7494c8aa72c220391c99800910856
+size 243

MultilingualG2PDecoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fb5b5003066a66bde1b5bff980e3267453910b04f96ca7352c0adc996ae7df8f
+size 507

MultilingualG2PDecoder.mlmodelc/metadata.json

@@ -0,0 +1,113 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float32",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "logits",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Range1d" : 2,
+      "Fill" : 2,
+      "Ios17.reshape" : 32,
+      "Ios16.reduceMean" : 13,
+      "Ios16.softmax" : 8,
+      "Ios17.matmul" : 16,
+      "Ios17.transpose" : 33,
+      "Select" : 3,
+      "Ios17.expandDims" : 11,
+      "Ios17.add" : 38,
+      "Tile" : 1,
+      "Ios17.lessEqual" : 1,
+      "Ios16.fillLike" : 1,
+      "Shape" : 4,
+      "Ios17.gather" : 6,
+      "Ios17.log" : 1,
+      "Ios17.cast" : 4,
+      "Ios17.sub" : 4,
+      "Ios17.pow" : 13,
+      "Ios17.less" : 1,
+      "Ios16.gelu" : 4,
+      "Ios17.linear" : 45,
+      "Ios17.concat" : 3,
+      "Ios17.greaterEqual" : 2,
+      "Ios17.minimum" : 2,
+      "Ios17.rsqrt" : 13,
+      "Ios17.mul" : 37
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-03-13",
+      "com.github.apple.coremltools.source" : "torch==2.7.0",
+      "com.github.apple.coremltools.version" : "9.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "inputSchema" : [
+      {
+        "dataType" : "Int32",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 1...128",
+        "shapeRange" : "[[1, 1], [1, 128]]",
+        "formattedType" : "MultiArray (Int32 1 × 8)",
+        "type" : "MultiArray",
+        "shape" : "[1, 8]",
+        "name" : "decoder_input_ids",
+        "shortDescription" : ""
+      },
+      {
+        "dataType" : "Float32",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 1...64 × 256",
+        "shapeRange" : "[[1, 1], [1, 64], [256, 256]]",
+        "formattedType" : "MultiArray (Float32 1 × 16 × 256)",
+        "type" : "MultiArray",
+        "shape" : "[1, 16, 256]",
+        "name" : "encoder_hidden_states",
+        "shortDescription" : ""
+      },
+      {
+        "dataType" : "Int32",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 1...64",
+        "shapeRange" : "[[1, 1], [1, 64]]",
+        "formattedType" : "MultiArray (Int32 1 × 16)",
+        "type" : "MultiArray",
+        "shape" : "[1, 16]",
+        "name" : "encoder_attention_mask",
+        "shortDescription" : ""
+      }
+    ],
+    "generatedClassName" : "G2PDecoder",
+    "method" : "predict"
+  }
+]

MultilingualG2PDecoder.mlmodelc/model.mil

@@ -0,0 +1,603 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.7.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios17>(tensor<int32, [1, ?]> decoder_input_ids, tensor<int32, [1, ?]> encoder_attention_mask, tensor<fp32, [1, ?, 256]> encoder_hidden_states) [FlexibleShapeInformation = tuple<tuple<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>, tuple<tensor<string, []>, dict<tensor<string, []>, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"decoder_input_ids", [1, 8]}, {"encoder_attention_mask", [1, 16]}, {"encoder_hidden_states", [1, 16, 256]}}), ("RangeDims", {{"decoder_input_ids", [[1, 1], [1, 128]]}, {"encoder_attention_mask", [[1, 1], [1, 64]]}, {"encoder_hidden_states", [[1, 1], [1, 64], [256, 256]]}})))] {
+            tensor<fp32, [384, 256]> decoder_embed_tokens_weight = const()[name = tensor<string, []>("decoder_embed_tokens_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp32, [256]> decoder_block_0_layer_0_layer_norm_weight = const()[name = tensor<string, []>("decoder_block_0_layer_0_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(393344)))];
+            tensor<fp32, [384, 256]> decoder_block_0_layer_0_SelfAttention_q_weight = const()[name = tensor<string, []>("decoder_block_0_layer_0_SelfAttention_q_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(394432)))];
+            tensor<fp32, [384, 256]> decoder_block_0_layer_0_SelfAttention_k_weight = const()[name = tensor<string, []>("decoder_block_0_layer_0_SelfAttention_k_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(787712)))];
+            tensor<fp32, [384, 256]> decoder_block_0_layer_0_SelfAttention_v_weight = const()[name = tensor<string, []>("decoder_block_0_layer_0_SelfAttention_v_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1180992)))];
+            tensor<fp32, [32, 6]> decoder_block_0_layer_0_SelfAttention_relative_attention_bias_weight = const()[name = tensor<string, []>("decoder_block_0_layer_0_SelfAttention_relative_attention_bias_weight"), val = tensor<fp32, [32, 6]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1574272)))];
+            tensor<fp32, [256, 384]> decoder_block_0_layer_0_SelfAttention_o_weight = const()[name = tensor<string, []>("decoder_block_0_layer_0_SelfAttention_o_weight"), val = tensor<fp32, [256, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1575104)))];
+            tensor<fp32, [256]> decoder_block_0_layer_1_layer_norm_weight = const()[name = tensor<string, []>("decoder_block_0_layer_1_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1968384)))];
+            tensor<fp32, [384, 256]> decoder_block_0_layer_1_EncDecAttention_q_weight = const()[name = tensor<string, []>("decoder_block_0_layer_1_EncDecAttention_q_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1969472)))];
+            tensor<fp32, [384, 256]> decoder_block_0_layer_1_EncDecAttention_k_weight = const()[name = tensor<string, []>("decoder_block_0_layer_1_EncDecAttention_k_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2362752)))];
+            tensor<fp32, [384, 256]> decoder_block_0_layer_1_EncDecAttention_v_weight = const()[name = tensor<string, []>("decoder_block_0_layer_1_EncDecAttention_v_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2756032)))];
+            tensor<fp32, [256, 384]> decoder_block_0_layer_1_EncDecAttention_o_weight = const()[name = tensor<string, []>("decoder_block_0_layer_1_EncDecAttention_o_weight"), val = tensor<fp32, [256, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3149312)))];
+            tensor<fp32, [256]> decoder_block_0_layer_2_layer_norm_weight = const()[name = tensor<string, []>("decoder_block_0_layer_2_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3542592)))];
+            tensor<fp32, [1024, 256]> decoder_block_0_layer_2_DenseReluDense_wi_0_weight = const()[name = tensor<string, []>("decoder_block_0_layer_2_DenseReluDense_wi_0_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3543680)))];
+            tensor<fp32, [1024, 256]> decoder_block_0_layer_2_DenseReluDense_wi_1_weight = const()[name = tensor<string, []>("decoder_block_0_layer_2_DenseReluDense_wi_1_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4592320)))];
+            tensor<fp32, [256, 1024]> decoder_block_0_layer_2_DenseReluDense_wo_weight = const()[name = tensor<string, []>("decoder_block_0_layer_2_DenseReluDense_wo_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5640960)))];
+            tensor<fp32, [256]> decoder_block_1_layer_0_layer_norm_weight = const()[name = tensor<string, []>("decoder_block_1_layer_0_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6689600)))];
+            tensor<fp32, [384, 256]> decoder_block_1_layer_0_SelfAttention_q_weight = const()[name = tensor<string, []>("decoder_block_1_layer_0_SelfAttention_q_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6690688)))];
+            tensor<fp32, [384, 256]> decoder_block_1_layer_0_SelfAttention_k_weight = const()[name = tensor<string, []>("decoder_block_1_layer_0_SelfAttention_k_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7083968)))];
+            tensor<fp32, [384, 256]> decoder_block_1_layer_0_SelfAttention_v_weight = const()[name = tensor<string, []>("decoder_block_1_layer_0_SelfAttention_v_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7477248)))];
+            tensor<fp32, [256, 384]> decoder_block_1_layer_0_SelfAttention_o_weight = const()[name = tensor<string, []>("decoder_block_1_layer_0_SelfAttention_o_weight"), val = tensor<fp32, [256, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7870528)))];
+            tensor<fp32, [256]> decoder_block_1_layer_1_layer_norm_weight = const()[name = tensor<string, []>("decoder_block_1_layer_1_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8263808)))];
+            tensor<fp32, [384, 256]> decoder_block_1_layer_1_EncDecAttention_q_weight = const()[name = tensor<string, []>("decoder_block_1_layer_1_EncDecAttention_q_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8264896)))];
+            tensor<fp32, [384, 256]> decoder_block_1_layer_1_EncDecAttention_k_weight = const()[name = tensor<string, []>("decoder_block_1_layer_1_EncDecAttention_k_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8658176)))];
+            tensor<fp32, [384, 256]> decoder_block_1_layer_1_EncDecAttention_v_weight = const()[name = tensor<string, []>("decoder_block_1_layer_1_EncDecAttention_v_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9051456)))];
+            tensor<fp32, [256, 384]> decoder_block_1_layer_1_EncDecAttention_o_weight = const()[name = tensor<string, []>("decoder_block_1_layer_1_EncDecAttention_o_weight"), val = tensor<fp32, [256, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9444736)))];
+            tensor<fp32, [256]> decoder_block_1_layer_2_layer_norm_weight = const()[name = tensor<string, []>("decoder_block_1_layer_2_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9838016)))];
+            tensor<fp32, [1024, 256]> decoder_block_1_layer_2_DenseReluDense_wi_0_weight = const()[name = tensor<string, []>("decoder_block_1_layer_2_DenseReluDense_wi_0_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9839104)))];
+            tensor<fp32, [1024, 256]> decoder_block_1_layer_2_DenseReluDense_wi_1_weight = const()[name = tensor<string, []>("decoder_block_1_layer_2_DenseReluDense_wi_1_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10887744)))];
+            tensor<fp32, [256, 1024]> decoder_block_1_layer_2_DenseReluDense_wo_weight = const()[name = tensor<string, []>("decoder_block_1_layer_2_DenseReluDense_wo_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11936384)))];
+            tensor<fp32, [256]> decoder_block_2_layer_0_layer_norm_weight = const()[name = tensor<string, []>("decoder_block_2_layer_0_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12985024)))];
+            tensor<fp32, [384, 256]> decoder_block_2_layer_0_SelfAttention_q_weight = const()[name = tensor<string, []>("decoder_block_2_layer_0_SelfAttention_q_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12986112)))];
+            tensor<fp32, [384, 256]> decoder_block_2_layer_0_SelfAttention_k_weight = const()[name = tensor<string, []>("decoder_block_2_layer_0_SelfAttention_k_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13379392)))];
+            tensor<fp32, [384, 256]> decoder_block_2_layer_0_SelfAttention_v_weight = const()[name = tensor<string, []>("decoder_block_2_layer_0_SelfAttention_v_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13772672)))];
+            tensor<fp32, [256, 384]> decoder_block_2_layer_0_SelfAttention_o_weight = const()[name = tensor<string, []>("decoder_block_2_layer_0_SelfAttention_o_weight"), val = tensor<fp32, [256, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14165952)))];
+            tensor<fp32, [256]> decoder_block_2_layer_1_layer_norm_weight = const()[name = tensor<string, []>("decoder_block_2_layer_1_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14559232)))];
+            tensor<fp32, [384, 256]> decoder_block_2_layer_1_EncDecAttention_q_weight = const()[name = tensor<string, []>("decoder_block_2_layer_1_EncDecAttention_q_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14560320)))];
+            tensor<fp32, [384, 256]> decoder_block_2_layer_1_EncDecAttention_k_weight = const()[name = tensor<string, []>("decoder_block_2_layer_1_EncDecAttention_k_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14953600)))];
+            tensor<fp32, [384, 256]> decoder_block_2_layer_1_EncDecAttention_v_weight = const()[name = tensor<string, []>("decoder_block_2_layer_1_EncDecAttention_v_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15346880)))];
+            tensor<fp32, [256, 384]> decoder_block_2_layer_1_EncDecAttention_o_weight = const()[name = tensor<string, []>("decoder_block_2_layer_1_EncDecAttention_o_weight"), val = tensor<fp32, [256, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15740160)))];
+            tensor<fp32, [256]> decoder_block_2_layer_2_layer_norm_weight = const()[name = tensor<string, []>("decoder_block_2_layer_2_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16133440)))];
+            tensor<fp32, [1024, 256]> decoder_block_2_layer_2_DenseReluDense_wi_0_weight = const()[name = tensor<string, []>("decoder_block_2_layer_2_DenseReluDense_wi_0_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16134528)))];
+            tensor<fp32, [1024, 256]> decoder_block_2_layer_2_DenseReluDense_wi_1_weight = const()[name = tensor<string, []>("decoder_block_2_layer_2_DenseReluDense_wi_1_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17183168)))];
+            tensor<fp32, [256, 1024]> decoder_block_2_layer_2_DenseReluDense_wo_weight = const()[name = tensor<string, []>("decoder_block_2_layer_2_DenseReluDense_wo_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18231808)))];
+            tensor<fp32, [256]> decoder_block_3_layer_0_layer_norm_weight = const()[name = tensor<string, []>("decoder_block_3_layer_0_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19280448)))];
+            tensor<fp32, [384, 256]> decoder_block_3_layer_0_SelfAttention_q_weight = const()[name = tensor<string, []>("decoder_block_3_layer_0_SelfAttention_q_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19281536)))];
+            tensor<fp32, [384, 256]> decoder_block_3_layer_0_SelfAttention_k_weight = const()[name = tensor<string, []>("decoder_block_3_layer_0_SelfAttention_k_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19674816)))];
+            tensor<fp32, [384, 256]> decoder_block_3_layer_0_SelfAttention_v_weight = const()[name = tensor<string, []>("decoder_block_3_layer_0_SelfAttention_v_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20068096)))];
+            tensor<fp32, [256, 384]> decoder_block_3_layer_0_SelfAttention_o_weight = const()[name = tensor<string, []>("decoder_block_3_layer_0_SelfAttention_o_weight"), val = tensor<fp32, [256, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20461376)))];
+            tensor<fp32, [256]> decoder_block_3_layer_1_layer_norm_weight = const()[name = tensor<string, []>("decoder_block_3_layer_1_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20854656)))];
+            tensor<fp32, [384, 256]> decoder_block_3_layer_1_EncDecAttention_q_weight = const()[name = tensor<string, []>("decoder_block_3_layer_1_EncDecAttention_q_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20855744)))];
+            tensor<fp32, [384, 256]> decoder_block_3_layer_1_EncDecAttention_k_weight = const()[name = tensor<string, []>("decoder_block_3_layer_1_EncDecAttention_k_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21249024)))];
+            tensor<fp32, [384, 256]> decoder_block_3_layer_1_EncDecAttention_v_weight = const()[name = tensor<string, []>("decoder_block_3_layer_1_EncDecAttention_v_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21642304)))];
+            tensor<fp32, [256, 384]> decoder_block_3_layer_1_EncDecAttention_o_weight = const()[name = tensor<string, []>("decoder_block_3_layer_1_EncDecAttention_o_weight"), val = tensor<fp32, [256, 384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22035584)))];
+            tensor<fp32, [256]> decoder_block_3_layer_2_layer_norm_weight = const()[name = tensor<string, []>("decoder_block_3_layer_2_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22428864)))];
+            tensor<fp32, [1024, 256]> decoder_block_3_layer_2_DenseReluDense_wi_0_weight = const()[name = tensor<string, []>("decoder_block_3_layer_2_DenseReluDense_wi_0_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22429952)))];
+            tensor<fp32, [1024, 256]> decoder_block_3_layer_2_DenseReluDense_wi_1_weight = const()[name = tensor<string, []>("decoder_block_3_layer_2_DenseReluDense_wi_1_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23478592)))];
+            tensor<fp32, [256, 1024]> decoder_block_3_layer_2_DenseReluDense_wo_weight = const()[name = tensor<string, []>("decoder_block_3_layer_2_DenseReluDense_wo_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24527232)))];
+            tensor<fp32, [256]> decoder_final_layer_norm_weight = const()[name = tensor<string, []>("decoder_final_layer_norm_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25575872)))];
+            tensor<fp32, [384, 256]> lm_head_weight = const()[name = tensor<string, []>("lm_head_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25576960)))];
+            tensor<int32, []> var_8 = const()[name = tensor<string, []>("op_8"), val = tensor<int32, []>(16)];
+            tensor<fp32, []> var_14 = const()[name = tensor<string, []>("op_14"), val = tensor<fp32, []>(0x1p+0)];
+            tensor<int32, []> var_16 = const()[name = tensor<string, []>("op_16"), val = tensor<int32, []>(6)];
+            tensor<int32, []> var_22 = const()[name = tensor<string, []>("op_22"), val = tensor<int32, []>(-1)];
+            tensor<int32, []> var_23 = const()[name = tensor<string, []>("op_23"), val = tensor<int32, []>(1)];
+            tensor<int32, [2]> var_35_shape = shape(x = decoder_input_ids)[name = tensor<string, []>("op_35_shape")];
+            tensor<int32, []> gather_0 = const()[name = tensor<string, []>("gather_0"), val = tensor<int32, []>(1)];
+            tensor<int32, []> gather_1_batch_dims_0 = const()[name = tensor<string, []>("gather_1_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_1_validate_indices_0 = const()[name = tensor<string, []>("gather_1_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<int32, []> select_0 = const()[name = tensor<string, []>("select_0"), val = tensor<int32, []>(1)];
+            tensor<int32, []> gather_1_axis_1 = const()[name = tensor<string, []>("gather_1_axis_1"), val = tensor<int32, []>(0)];
+            tensor<int32, []> gather_1 = gather(axis = gather_1_axis_1, batch_dims = gather_1_batch_dims_0, indices = select_0, validate_indices = gather_1_validate_indices_0, x = var_35_shape)[name = tensor<string, []>("gather_1")];
+            tensor<int32, []> input_3_batch_dims_0 = const()[name = tensor<string, []>("input_3_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> input_3_validate_indices_0 = const()[name = tensor<string, []>("input_3_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<int32, []> greater_equal_1_y_0 = const()[name = tensor<string, []>("greater_equal_1_y_0"), val = tensor<int32, []>(0)];
+            tensor<bool, [1, ?]> greater_equal_1 = greater_equal(x = decoder_input_ids, y = greater_equal_1_y_0)[name = tensor<string, []>("greater_equal_1")];
+            tensor<int32, []> slice_by_index_1 = const()[name = tensor<string, []>("slice_by_index_1"), val = tensor<int32, []>(384)];
+            tensor<int32, [1, ?]> add_1 = add(x = decoder_input_ids, y = slice_by_index_1)[name = tensor<string, []>("add_1")];
+            tensor<int32, [1, ?]> select_1 = select(a = decoder_input_ids, b = add_1, cond = greater_equal_1)[name = tensor<string, []>("select_1")];
+            tensor<int32, []> input_3_axis_1 = const()[name = tensor<string, []>("input_3_axis_1"), val = tensor<int32, []>(0)];
+            tensor<fp32, [1, ?, 256]> input_3 = gather(axis = input_3_axis_1, batch_dims = input_3_batch_dims_0, indices = select_1, validate_indices = input_3_validate_indices_0, x = decoder_embed_tokens_weight)[name = tensor<string, []>("input_3")];
+            tensor<int32, []> concat_1_axis_0 = const()[name = tensor<string, []>("concat_1_axis_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> concat_1_interleave_0 = const()[name = tensor<string, []>("concat_1_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [2]> concat_1 = concat(axis = concat_1_axis_0, interleave = concat_1_interleave_0, values = (gather_0, gather_1))[name = tensor<string, []>("concat_1")];
+            tensor<fp32, []> fill_0_value_0 = const()[name = tensor<string, []>("fill_0_value_0"), val = tensor<fp32, []>(0x1p+0)];
+            tensor<fp32, [1, ?]> fill_0 = fill(shape = concat_1, value = fill_0_value_0)[name = tensor<string, []>("fill_0")];
+            tensor<int32, []> const_0 = const()[name = tensor<string, []>("const_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> const_1 = const()[name = tensor<string, []>("const_1"), val = tensor<int32, []>(1)];
+            tensor<int32, [?]> seq_ids = range_1d(end = gather_1, start = const_0, step = const_1)[name = tensor<string, []>("seq_ids")];
+            tensor<int32, [1]> var_44_axes_0 = const()[name = tensor<string, []>("op_44_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> var_44 = expand_dims(axes = var_44_axes_0, x = seq_ids)[name = tensor<string, []>("op_44")];
+            tensor<int32, [1]> var_45_axes_0 = const()[name = tensor<string, []>("op_45_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1, ?]> var_45 = expand_dims(axes = var_45_axes_0, x = var_44)[name = tensor<string, []>("op_45")];
+            tensor<int32, []> concat_2_axis_0 = const()[name = tensor<string, []>("concat_2_axis_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> concat_2_interleave_0 = const()[name = tensor<string, []>("concat_2_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [3]> concat_2 = concat(axis = concat_2_axis_0, interleave = concat_2_interleave_0, values = (gather_0, gather_1, var_23))[name = tensor<string, []>("concat_2")];
+            tensor<int32, [1, ?, ?]> var_48 = tile(reps = concat_2, x = var_45)[name = tensor<string, []>("op_48")];
+            tensor<int32, [1]> var_51_axes_0 = const()[name = tensor<string, []>("op_51_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<int32, [1, ?, 1]> var_51 = expand_dims(axes = var_51_axes_0, x = var_44)[name = tensor<string, []>("op_51")];
+            tensor<bool, [1, ?, ?]> causal_mask_1 = less_equal(x = var_48, y = var_51)[name = tensor<string, []>("causal_mask_1")];
+            tensor<string, []> causal_mask_dtype_0 = const()[name = tensor<string, []>("causal_mask_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> var_55_axes_0 = const()[name = tensor<string, []>("op_55_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, ?, ?]> causal_mask = cast(dtype = causal_mask_dtype_0, x = causal_mask_1)[name = tensor<string, []>("cast_42")];
+            tensor<fp32, [1, 1, ?, ?]> var_55 = expand_dims(axes = var_55_axes_0, x = causal_mask)[name = tensor<string, []>("op_55")];
+            tensor<int32, [1]> var_59_axes_0 = const()[name = tensor<string, []>("op_59_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1, ?]> var_59 = expand_dims(axes = var_59_axes_0, x = fill_0)[name = tensor<string, []>("op_59")];
+            tensor<int32, [1]> var_60_axes_0 = const()[name = tensor<string, []>("op_60_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp32, [1, 1, 1, ?]> var_60 = expand_dims(axes = var_60_axes_0, x = var_59)[name = tensor<string, []>("op_60")];
+            tensor<fp32, [1, 1, ?, ?]> extended_attention_mask = mul(x = var_55, y = var_60)[name = tensor<string, []>("extended_attention_mask")];
+            tensor<fp32, [1, 1, ?, ?]> var_64 = sub(x = var_14, y = extended_attention_mask)[name = tensor<string, []>("op_64")];
+            tensor<fp32, []> var_65 = const()[name = tensor<string, []>("op_65"), val = tensor<fp32, []>(-0x1.fffffep+127)];
+            tensor<fp32, [1, 1, ?, ?]> mask_1 = mul(x = var_64, y = var_65)[name = tensor<string, []>("mask_1")];
+            tensor<int32, [1]> var_68_axes_0 = const()[name = tensor<string, []>("op_68_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1, ?]> var_68 = expand_dims(axes = var_68_axes_0, x = encoder_attention_mask)[name = tensor<string, []>("op_68")];
+            tensor<int32, [1]> var_69_axes_0 = const()[name = tensor<string, []>("op_69_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<int32, [1, 1, 1, ?]> var_69 = expand_dims(axes = var_69_axes_0, x = var_68)[name = tensor<string, []>("op_69")];
+            tensor<string, []> var_71_dtype_0 = const()[name = tensor<string, []>("op_71_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, [1, 1, 1, ?]> var_71 = cast(dtype = var_71_dtype_0, x = var_69)[name = tensor<string, []>("cast_41")];
+            tensor<fp32, [1, 1, 1, ?]> var_72 = sub(x = var_14, y = var_71)[name = tensor<string, []>("op_72")];
+            tensor<fp32, []> var_73 = const()[name = tensor<string, []>("op_73"), val = tensor<fp32, []>(-0x1.fffffep+127)];
+            tensor<fp32, [1, 1, 1, ?]> mask = mul(x = var_72, y = var_73)[name = tensor<string, []>("mask")];
+            tensor<fp32, []> var_18_promoted = const()[name = tensor<string, []>("op_18_promoted"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1, ?, 256]> var_86 = pow(x = input_3, y = var_18_promoted)[name = tensor<string, []>("op_86")];
+            tensor<int32, [1]> variance_1_axes_0 = const()[name = tensor<string, []>("variance_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> variance_1_keep_dims_0 = const()[name = tensor<string, []>("variance_1_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp32, [1, ?, 1]> variance_1 = reduce_mean(axes = variance_1_axes_0, keep_dims = variance_1_keep_dims_0, x = var_86)[name = tensor<string, []>("variance_1")];
+            tensor<fp32, []> var_89 = const()[name = tensor<string, []>("op_89"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
+            tensor<fp32, [1, ?, 1]> var_90 = add(x = variance_1, y = var_89)[name = tensor<string, []>("op_90")];
+            tensor<fp32, []> var_91_epsilon_0 = const()[name = tensor<string, []>("op_91_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp32, [1, ?, 1]> var_91 = rsqrt(epsilon = var_91_epsilon_0, x = var_90)[name = tensor<string, []>("op_91")];
+            tensor<fp32, [1, ?, 256]> hidden_states_5 = mul(x = input_3, y = var_91)[name = tensor<string, []>("hidden_states_5")];
+            tensor<fp32, [1, ?, 256]> hidden_states_7 = mul(x = decoder_block_0_layer_0_layer_norm_weight, y = hidden_states_5)[name = tensor<string, []>("hidden_states_7")];
+            tensor<int32, [3]> var_103_shape = shape(x = hidden_states_7)[name = tensor<string, []>("op_103_shape")];
+            tensor<int32, []> gather_3_batch_dims_0 = const()[name = tensor<string, []>("gather_3_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_3_validate_indices_0 = const()[name = tensor<string, []>("gather_3_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<int32, []> select_2 = const()[name = tensor<string, []>("select_2"), val = tensor<int32, []>(1)];
+            tensor<int32, []> gather_3_axis_1 = const()[name = tensor<string, []>("gather_3_axis_1"), val = tensor<int32, []>(0)];
+            tensor<int32, []> gather_3 = gather(axis = gather_3_axis_1, batch_dims = gather_3_batch_dims_0, indices = select_2, validate_indices = gather_3_validate_indices_0, x = var_103_shape)[name = tensor<string, []>("gather_3")];
+            tensor<fp32, [384]> linear_0_bias_0 = const()[name = tensor<string, []>("linear_0_bias_0"), val = tensor<fp32, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25970240)))];
+            tensor<fp32, [1, ?, 384]> states_1 = linear(bias = linear_0_bias_0, weight = decoder_block_0_layer_0_SelfAttention_q_weight, x = hidden_states_7)[name = tensor<string, []>("linear_0")];
+            tensor<int32, [4]> var_106 = const()[name = tensor<string, []>("op_106"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_107 = reshape(shape = var_106, x = states_1)[name = tensor<string, []>("op_107")];
+            tensor<fp32, [1, ?, 384]> states_3 = linear(bias = linear_0_bias_0, weight = decoder_block_0_layer_0_SelfAttention_k_weight, x = hidden_states_7)[name = tensor<string, []>("linear_1")];
+            tensor<int32, [4]> var_111 = const()[name = tensor<string, []>("op_111"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_112 = reshape(shape = var_111, x = states_3)[name = tensor<string, []>("op_112")];
+            tensor<fp32, [1, ?, 384]> states_5 = linear(bias = linear_0_bias_0, weight = decoder_block_0_layer_0_SelfAttention_v_weight, x = hidden_states_7)[name = tensor<string, []>("linear_2")];
+            tensor<int32, [4]> var_116 = const()[name = tensor<string, []>("op_116"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_117 = reshape(shape = var_116, x = states_5)[name = tensor<string, []>("op_117")];
+            tensor<int32, [4]> value_states_1_perm_0 = const()[name = tensor<string, []>("value_states_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> scores_1_transpose_x_0 = const()[name = tensor<string, []>("scores_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> scores_1_transpose_y_0 = const()[name = tensor<string, []>("scores_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_24_perm_0 = const()[name = tensor<string, []>("transpose_24_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_25_perm_0 = const()[name = tensor<string, []>("transpose_25_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 6, 64, ?]> transpose_25 = transpose(perm = transpose_25_perm_0, x = var_112)[name = tensor<string, []>("transpose_70")];
+            tensor<fp32, [1, 6, ?, 64]> transpose_24 = transpose(perm = transpose_24_perm_0, x = var_107)[name = tensor<string, []>("transpose_71")];
+            tensor<fp32, [1, 6, ?, ?]> scores_1 = matmul(transpose_x = scores_1_transpose_x_0, transpose_y = scores_1_transpose_y_0, x = transpose_24, y = transpose_25)[name = tensor<string, []>("scores_1")];
+            tensor<int32, []> const_2 = const()[name = tensor<string, []>("const_2"), val = tensor<int32, []>(0)];
+            tensor<int32, []> const_3 = const()[name = tensor<string, []>("const_3"), val = tensor<int32, []>(1)];
+            tensor<int32, [?]> var_121 = range_1d(end = gather_3, start = const_2, step = const_3)[name = tensor<string, []>("op_121")];
+            tensor<int32, [1]> context_position_axes_0 = const()[name = tensor<string, []>("context_position_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [?, 1]> context_position = expand_dims(axes = context_position_axes_0, x = var_121)[name = tensor<string, []>("context_position")];
+            tensor<int32, [1]> var_125_axes_0 = const()[name = tensor<string, []>("op_125_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> var_125 = expand_dims(axes = var_125_axes_0, x = var_121)[name = tensor<string, []>("op_125")];
+            tensor<int32, [?, ?]> relative_position_1 = sub(x = var_125, y = context_position)[name = tensor<string, []>("relative_position_1")];
+            tensor<int32, [?, ?]> var_128 = sub(x = relative_position_1, y = relative_position_1)[name = tensor<string, []>("sub_0")];
+            tensor<int32, [?, ?]> var_129 = minimum(x = relative_position_1, y = var_128)[name = tensor<string, []>("op_129")];
+            tensor<int32, []> const_6 = const()[name = tensor<string, []>("const_6"), val = tensor<int32, []>(-1)];
+            tensor<int32, [?, ?]> relative_position = mul(x = var_129, y = const_6)[name = tensor<string, []>("relative_position")];
+            tensor<bool, [?, ?]> is_small = less(x = relative_position, y = var_8)[name = tensor<string, []>("is_small")];
+            tensor<string, []> var_132_dtype_0 = const()[name = tensor<string, []>("op_132_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, []> _inversed_134_y_0 = const()[name = tensor<string, []>("_inversed_134_y_0"), val = tensor<fp32, []>(0x1p-4)];
+            tensor<fp32, [?, ?]> var_132 = cast(dtype = var_132_dtype_0, x = relative_position)[name = tensor<string, []>("cast_40")];
+            tensor<fp32, [?, ?]> _inversed_134 = mul(x = var_132, y = _inversed_134_y_0)[name = tensor<string, []>("_inversed_134")];
+            tensor<fp32, []> var_135_epsilon_0 = const()[name = tensor<string, []>("op_135_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
+            tensor<fp32, [?, ?]> var_135 = log(epsilon = var_135_epsilon_0, x = _inversed_134)[name = tensor<string, []>("op_135")];
+            tensor<fp32, []> _inversed_137_y_0 = const()[name = tensor<string, []>("_inversed_137_y_0"), val = tensor<fp32, []>(0x1.ec709ep-2)];
+            tensor<fp32, [?, ?]> _inversed_137 = mul(x = var_135, y = _inversed_137_y_0)[name = tensor<string, []>("_inversed_137")];
+            tensor<fp32, []> var_138_promoted = const()[name = tensor<string, []>("op_138_promoted"), val = tensor<fp32, []>(0x1p+4)];
+            tensor<fp32, [?, ?]> var_139 = mul(x = _inversed_137, y = var_138_promoted)[name = tensor<string, []>("op_139")];
+            tensor<string, []> var_140_dtype_0 = const()[name = tensor<string, []>("op_140_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, []> var_141 = const()[name = tensor<string, []>("op_141"), val = tensor<int32, []>(16)];
+            tensor<int32, [?, ?]> var_140 = cast(dtype = var_140_dtype_0, x = var_139)[name = tensor<string, []>("cast_39")];
+            tensor<int32, [?, ?]> relative_position_if_large_1 = add(x = var_140, y = var_141)[name = tensor<string, []>("relative_position_if_large_1")];
+            tensor<int32, []> var_143_value_0 = const()[name = tensor<string, []>("op_143_value_0"), val = tensor<int32, []>(31)];
+            tensor<int32, [?, ?]> var_143 = fill_like(ref_tensor = relative_position_if_large_1, value = var_143_value_0)[name = tensor<string, []>("op_143")];
+            tensor<int32, [?, ?]> relative_position_if_large = minimum(x = relative_position_if_large_1, y = var_143)[name = tensor<string, []>("relative_position_if_large")];
+            tensor<int32, [?, ?]> var_145 = select(a = relative_position, b = relative_position_if_large, cond = is_small)[name = tensor<string, []>("op_145")];
+            tensor<int32, []> values_batch_dims_0 = const()[name = tensor<string, []>("values_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> values_validate_indices_0 = const()[name = tensor<string, []>("values_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<int32, []> greater_equal_3_y_0 = const()[name = tensor<string, []>("greater_equal_3_y_0"), val = tensor<int32, []>(0)];
+            tensor<bool, [?, ?]> greater_equal_3 = greater_equal(x = var_145, y = greater_equal_3_y_0)[name = tensor<string, []>("greater_equal_3")];
+            tensor<int32, []> slice_by_index_3 = const()[name = tensor<string, []>("slice_by_index_3"), val = tensor<int32, []>(32)];
+            tensor<int32, [?, ?]> add_3 = add(x = var_145, y = slice_by_index_3)[name = tensor<string, []>("add_3")];
+            tensor<int32, [?, ?]> select_3 = select(a = var_145, b = add_3, cond = greater_equal_3)[name = tensor<string, []>("select_3")];
+            tensor<int32, []> values_axis_1 = const()[name = tensor<string, []>("values_axis_1"), val = tensor<int32, []>(0)];
+            tensor<fp32, [?, ?, 6]> values = gather(axis = values_axis_1, batch_dims = values_batch_dims_0, indices = select_3, validate_indices = values_validate_indices_0, x = decoder_block_0_layer_0_SelfAttention_relative_attention_bias_weight)[name = tensor<string, []>("values")];
+            tensor<int32, [3]> var_150 = const()[name = tensor<string, []>("op_150"), val = tensor<int32, [3]>([2, 0, 1])];
+            tensor<int32, [1]> position_bias_1_axes_0 = const()[name = tensor<string, []>("position_bias_1_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [6, ?, ?]> var_151 = transpose(perm = var_150, x = values)[name = tensor<string, []>("transpose_69")];
+            tensor<fp32, [1, 6, ?, ?]> position_bias_1 = expand_dims(axes = position_bias_1_axes_0, x = var_151)[name = tensor<string, []>("position_bias_1")];
+            tensor<fp32, [1, 6, ?, ?]> position_bias_3 = add(x = position_bias_1, y = mask_1)[name = tensor<string, []>("position_bias_3")];
+            tensor<fp32, [1, 6, ?, ?]> scores_3 = add(x = scores_1, y = position_bias_3)[name = tensor<string, []>("scores_3")];
+            tensor<fp32, [1, 6, ?, ?]> var_156 = softmax(axis = var_22, x = scores_3)[name = tensor<string, []>("op_156")];
+            tensor<bool, []> states_7_transpose_x_0 = const()[name = tensor<string, []>("states_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> states_7_transpose_y_0 = const()[name = tensor<string, []>("states_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 6, ?, 64]> value_states_1 = transpose(perm = value_states_1_perm_0, x = var_117)[name = tensor<string, []>("transpose_72")];
+            tensor<fp32, [1, 6, ?, 64]> states_7 = matmul(transpose_x = states_7_transpose_x_0, transpose_y = states_7_transpose_y_0, x = var_156, y = value_states_1)[name = tensor<string, []>("states_7")];
+            tensor<int32, [4]> var_160_perm_0 = const()[name = tensor<string, []>("op_160_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_162 = const()[name = tensor<string, []>("op_162"), val = tensor<int32, [3]>([1, -1, 384])];
+            tensor<fp32, [1, ?, 6, 64]> var_160 = transpose(perm = var_160_perm_0, x = states_7)[name = tensor<string, []>("transpose_68")];
+            tensor<fp32, [1, ?, 384]> input_11 = reshape(shape = var_162, x = var_160)[name = tensor<string, []>("input_11")];
+            tensor<fp32, [256]> linear_3_bias_0 = const()[name = tensor<string, []>("linear_3_bias_0"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25971840)))];
+            tensor<fp32, [1, ?, 256]> input_13 = linear(bias = linear_3_bias_0, weight = decoder_block_0_layer_0_SelfAttention_o_weight, x = input_11)[name = tensor<string, []>("linear_3")];
+            tensor<fp32, [1, ?, 256]> hidden_states_9 = add(x = input_3, y = input_13)[name = tensor<string, []>("hidden_states_9")];
+            tensor<fp32, []> var_18_promoted_1 = const()[name = tensor<string, []>("op_18_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1, ?, 256]> var_180 = pow(x = hidden_states_9, y = var_18_promoted_1)[name = tensor<string, []>("op_180")];
+            tensor<int32, [1]> variance_3_axes_0 = const()[name = tensor<string, []>("variance_3_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> variance_3_keep_dims_0 = const()[name = tensor<string, []>("variance_3_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp32, [1, ?, 1]> variance_3 = reduce_mean(axes = variance_3_axes_0, keep_dims = variance_3_keep_dims_0, x = var_180)[name = tensor<string, []>("variance_3")];
+            tensor<fp32, []> var_183 = const()[name = tensor<string, []>("op_183"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
+            tensor<fp32, [1, ?, 1]> var_184 = add(x = variance_3, y = var_183)[name = tensor<string, []>("op_184")];
+            tensor<fp32, []> var_185_epsilon_0 = const()[name = tensor<string, []>("op_185_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp32, [1, ?, 1]> var_185 = rsqrt(epsilon = var_185_epsilon_0, x = var_184)[name = tensor<string, []>("op_185")];
+            tensor<fp32, [1, ?, 256]> hidden_states_13 = mul(x = hidden_states_9, y = var_185)[name = tensor<string, []>("hidden_states_13")];
+            tensor<fp32, [1, ?, 256]> hidden_states_15 = mul(x = decoder_block_0_layer_1_layer_norm_weight, y = hidden_states_13)[name = tensor<string, []>("hidden_states_15")];
+            tensor<int32, [3]> var_196_shape = shape(x = hidden_states_15)[name = tensor<string, []>("op_196_shape")];
+            tensor<int32, []> gather_5_batch_dims_0 = const()[name = tensor<string, []>("gather_5_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_5_validate_indices_0 = const()[name = tensor<string, []>("gather_5_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<int32, []> select_4 = const()[name = tensor<string, []>("select_4"), val = tensor<int32, []>(1)];
+            tensor<int32, []> gather_5_axis_1 = const()[name = tensor<string, []>("gather_5_axis_1"), val = tensor<int32, []>(0)];
+            tensor<int32, []> gather_5 = gather(axis = gather_5_axis_1, batch_dims = gather_5_batch_dims_0, indices = select_4, validate_indices = gather_5_validate_indices_0, x = var_196_shape)[name = tensor<string, []>("gather_5")];
+            tensor<int32, [3]> var_197_shape = shape(x = encoder_hidden_states)[name = tensor<string, []>("op_197_shape")];
+            tensor<int32, []> gather_6_batch_dims_0 = const()[name = tensor<string, []>("gather_6_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_6_validate_indices_0 = const()[name = tensor<string, []>("gather_6_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<int32, []> select_5 = const()[name = tensor<string, []>("select_5"), val = tensor<int32, []>(1)];
+            tensor<int32, []> gather_6_axis_1 = const()[name = tensor<string, []>("gather_6_axis_1"), val = tensor<int32, []>(0)];
+            tensor<int32, []> gather_6 = gather(axis = gather_6_axis_1, batch_dims = gather_6_batch_dims_0, indices = select_5, validate_indices = gather_6_validate_indices_0, x = var_197_shape)[name = tensor<string, []>("gather_6")];
+            tensor<fp32, [1, ?, 384]> states_9 = linear(bias = linear_0_bias_0, weight = decoder_block_0_layer_1_EncDecAttention_q_weight, x = hidden_states_15)[name = tensor<string, []>("linear_4")];
+            tensor<int32, [4]> var_200 = const()[name = tensor<string, []>("op_200"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_201 = reshape(shape = var_200, x = states_9)[name = tensor<string, []>("op_201")];
+            tensor<fp32, [1, ?, 384]> states_11 = linear(bias = linear_0_bias_0, weight = decoder_block_0_layer_1_EncDecAttention_k_weight, x = encoder_hidden_states)[name = tensor<string, []>("linear_5")];
+            tensor<int32, [4]> var_205 = const()[name = tensor<string, []>("op_205"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_206 = reshape(shape = var_205, x = states_11)[name = tensor<string, []>("op_206")];
+            tensor<fp32, [1, ?, 384]> states_13 = linear(bias = linear_0_bias_0, weight = decoder_block_0_layer_1_EncDecAttention_v_weight, x = encoder_hidden_states)[name = tensor<string, []>("linear_6")];
+            tensor<int32, [4]> var_210 = const()[name = tensor<string, []>("op_210"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_211 = reshape(shape = var_210, x = states_13)[name = tensor<string, []>("op_211")];
+            tensor<int32, [4]> value_states_3_perm_0 = const()[name = tensor<string, []>("value_states_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> scores_5_transpose_x_0 = const()[name = tensor<string, []>("scores_5_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> scores_5_transpose_y_0 = const()[name = tensor<string, []>("scores_5_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_26_perm_0 = const()[name = tensor<string, []>("transpose_26_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_27_perm_0 = const()[name = tensor<string, []>("transpose_27_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 6, 64, ?]> transpose_27 = transpose(perm = transpose_27_perm_0, x = var_206)[name = tensor<string, []>("transpose_65")];
+            tensor<fp32, [1, 6, ?, 64]> transpose_26 = transpose(perm = transpose_26_perm_0, x = var_201)[name = tensor<string, []>("transpose_66")];
+            tensor<fp32, [1, 6, ?, ?]> scores_5 = matmul(transpose_x = scores_5_transpose_x_0, transpose_y = scores_5_transpose_y_0, x = transpose_26, y = transpose_27)[name = tensor<string, []>("scores_5")];
+            tensor<int32, []> concat_3_axis_0 = const()[name = tensor<string, []>("concat_3_axis_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> concat_3_interleave_0 = const()[name = tensor<string, []>("concat_3_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> concat_3 = concat(axis = concat_3_axis_0, interleave = concat_3_interleave_0, values = (var_23, var_16, gather_5, gather_6))[name = tensor<string, []>("concat_3")];
+            tensor<fp32, []> position_bias_5_value_0 = const()[name = tensor<string, []>("position_bias_5_value_0"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<fp32, [1, 6, ?, ?]> position_bias_5 = fill(shape = concat_3, value = position_bias_5_value_0)[name = tensor<string, []>("position_bias_5")];
+            tensor<fp32, [1, 6, ?, ?]> position_bias = add(x = position_bias_5, y = mask)[name = tensor<string, []>("position_bias")];
+            tensor<fp32, [1, 6, ?, ?]> scores_7 = add(x = scores_5, y = position_bias)[name = tensor<string, []>("scores_7")];
+            tensor<fp32, [1, 6, ?, ?]> var_220 = softmax(axis = var_22, x = scores_7)[name = tensor<string, []>("op_220")];
+            tensor<bool, []> states_15_transpose_x_0 = const()[name = tensor<string, []>("states_15_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> states_15_transpose_y_0 = const()[name = tensor<string, []>("states_15_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 6, ?, 64]> value_states_3 = transpose(perm = value_states_3_perm_0, x = var_211)[name = tensor<string, []>("transpose_67")];
+            tensor<fp32, [1, 6, ?, 64]> states_15 = matmul(transpose_x = states_15_transpose_x_0, transpose_y = states_15_transpose_y_0, x = var_220, y = value_states_3)[name = tensor<string, []>("states_15")];
+            tensor<int32, [4]> var_224_perm_0 = const()[name = tensor<string, []>("op_224_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_226 = const()[name = tensor<string, []>("op_226"), val = tensor<int32, [3]>([1, -1, 384])];
+            tensor<fp32, [1, ?, 6, 64]> var_224 = transpose(perm = var_224_perm_0, x = states_15)[name = tensor<string, []>("transpose_64")];
+            tensor<fp32, [1, ?, 384]> input_19 = reshape(shape = var_226, x = var_224)[name = tensor<string, []>("input_19")];
+            tensor<fp32, [1, ?, 256]> input_21 = linear(bias = linear_3_bias_0, weight = decoder_block_0_layer_1_EncDecAttention_o_weight, x = input_19)[name = tensor<string, []>("linear_7")];
+            tensor<fp32, [1, ?, 256]> hidden_states_17 = add(x = hidden_states_9, y = input_21)[name = tensor<string, []>("hidden_states_17")];
+            tensor<fp32, []> var_18_promoted_2 = const()[name = tensor<string, []>("op_18_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1, ?, 256]> var_242 = pow(x = hidden_states_17, y = var_18_promoted_2)[name = tensor<string, []>("op_242")];
+            tensor<int32, [1]> variance_5_axes_0 = const()[name = tensor<string, []>("variance_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> variance_5_keep_dims_0 = const()[name = tensor<string, []>("variance_5_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp32, [1, ?, 1]> variance_5 = reduce_mean(axes = variance_5_axes_0, keep_dims = variance_5_keep_dims_0, x = var_242)[name = tensor<string, []>("variance_5")];
+            tensor<fp32, []> var_245 = const()[name = tensor<string, []>("op_245"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
+            tensor<fp32, [1, ?, 1]> var_246 = add(x = variance_5, y = var_245)[name = tensor<string, []>("op_246")];
+            tensor<fp32, []> var_247_epsilon_0 = const()[name = tensor<string, []>("op_247_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp32, [1, ?, 1]> var_247 = rsqrt(epsilon = var_247_epsilon_0, x = var_246)[name = tensor<string, []>("op_247")];
+            tensor<fp32, [1, ?, 256]> hidden_states_21 = mul(x = hidden_states_17, y = var_247)[name = tensor<string, []>("hidden_states_21")];
+            tensor<fp32, [1, ?, 256]> input_23 = mul(x = decoder_block_0_layer_2_layer_norm_weight, y = hidden_states_21)[name = tensor<string, []>("input_23")];
+            tensor<fp32, [1024]> linear_8_bias_0 = const()[name = tensor<string, []>("linear_8_bias_0"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25972928)))];
+            tensor<fp32, [1, ?, 1024]> input_25 = linear(bias = linear_8_bias_0, weight = decoder_block_0_layer_2_DenseReluDense_wi_0_weight, x = input_23)[name = tensor<string, []>("linear_8")];
+            tensor<string, []> hidden_gelu_1_mode_0 = const()[name = tensor<string, []>("hidden_gelu_1_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
+            tensor<fp32, [1, ?, 1024]> hidden_gelu_1 = gelu(mode = hidden_gelu_1_mode_0, x = input_25)[name = tensor<string, []>("hidden_gelu_1")];
+            tensor<fp32, [1, ?, 1024]> hidden_linear_1 = linear(bias = linear_8_bias_0, weight = decoder_block_0_layer_2_DenseReluDense_wi_1_weight, x = input_23)[name = tensor<string, []>("linear_9")];
+            tensor<fp32, [1, ?, 1024]> input_27 = mul(x = hidden_gelu_1, y = hidden_linear_1)[name = tensor<string, []>("input_27")];
+            tensor<fp32, [1, ?, 256]> input_31 = linear(bias = linear_3_bias_0, weight = decoder_block_0_layer_2_DenseReluDense_wo_weight, x = input_27)[name = tensor<string, []>("linear_10")];
+            tensor<fp32, [1, ?, 256]> hidden_states_23 = add(x = hidden_states_17, y = input_31)[name = tensor<string, []>("hidden_states_23")];
+            tensor<fp32, []> var_18_promoted_3 = const()[name = tensor<string, []>("op_18_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1, ?, 256]> var_292 = pow(x = hidden_states_23, y = var_18_promoted_3)[name = tensor<string, []>("op_292")];
+            tensor<int32, [1]> variance_7_axes_0 = const()[name = tensor<string, []>("variance_7_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> variance_7_keep_dims_0 = const()[name = tensor<string, []>("variance_7_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp32, [1, ?, 1]> variance_7 = reduce_mean(axes = variance_7_axes_0, keep_dims = variance_7_keep_dims_0, x = var_292)[name = tensor<string, []>("variance_7")];
+            tensor<fp32, []> var_295 = const()[name = tensor<string, []>("op_295"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
+            tensor<fp32, [1, ?, 1]> var_296 = add(x = variance_7, y = var_295)[name = tensor<string, []>("op_296")];
+            tensor<fp32, []> var_297_epsilon_0 = const()[name = tensor<string, []>("op_297_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp32, [1, ?, 1]> var_297 = rsqrt(epsilon = var_297_epsilon_0, x = var_296)[name = tensor<string, []>("op_297")];
+            tensor<fp32, [1, ?, 256]> hidden_states_27 = mul(x = hidden_states_23, y = var_297)[name = tensor<string, []>("hidden_states_27")];
+            tensor<fp32, [1, ?, 256]> hidden_states_29 = mul(x = decoder_block_1_layer_0_layer_norm_weight, y = hidden_states_27)[name = tensor<string, []>("hidden_states_29")];
+            tensor<fp32, [1, ?, 384]> states_17 = linear(bias = linear_0_bias_0, weight = decoder_block_1_layer_0_SelfAttention_q_weight, x = hidden_states_29)[name = tensor<string, []>("linear_11")];
+            tensor<int32, [4]> var_310 = const()[name = tensor<string, []>("op_310"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_311 = reshape(shape = var_310, x = states_17)[name = tensor<string, []>("op_311")];
+            tensor<fp32, [1, ?, 384]> states_19 = linear(bias = linear_0_bias_0, weight = decoder_block_1_layer_0_SelfAttention_k_weight, x = hidden_states_29)[name = tensor<string, []>("linear_12")];
+            tensor<int32, [4]> var_315 = const()[name = tensor<string, []>("op_315"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_316 = reshape(shape = var_315, x = states_19)[name = tensor<string, []>("op_316")];
+            tensor<fp32, [1, ?, 384]> states_21 = linear(bias = linear_0_bias_0, weight = decoder_block_1_layer_0_SelfAttention_v_weight, x = hidden_states_29)[name = tensor<string, []>("linear_13")];
+            tensor<int32, [4]> var_320 = const()[name = tensor<string, []>("op_320"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_321 = reshape(shape = var_320, x = states_21)[name = tensor<string, []>("op_321")];
+            tensor<int32, [4]> value_states_5_perm_0 = const()[name = tensor<string, []>("value_states_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> scores_9_transpose_x_0 = const()[name = tensor<string, []>("scores_9_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> scores_9_transpose_y_0 = const()[name = tensor<string, []>("scores_9_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_28_perm_0 = const()[name = tensor<string, []>("transpose_28_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_29_perm_0 = const()[name = tensor<string, []>("transpose_29_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 6, 64, ?]> transpose_29 = transpose(perm = transpose_29_perm_0, x = var_316)[name = tensor<string, []>("transpose_61")];
+            tensor<fp32, [1, 6, ?, 64]> transpose_28 = transpose(perm = transpose_28_perm_0, x = var_311)[name = tensor<string, []>("transpose_62")];
+            tensor<fp32, [1, 6, ?, ?]> scores_9 = matmul(transpose_x = scores_9_transpose_x_0, transpose_y = scores_9_transpose_y_0, x = transpose_28, y = transpose_29)[name = tensor<string, []>("scores_9")];
+            tensor<fp32, [1, 6, ?, ?]> scores_11 = add(x = scores_9, y = position_bias_3)[name = tensor<string, []>("scores_11")];
+            tensor<fp32, [1, 6, ?, ?]> var_327 = softmax(axis = var_22, x = scores_11)[name = tensor<string, []>("op_327")];
+            tensor<bool, []> states_23_transpose_x_0 = const()[name = tensor<string, []>("states_23_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> states_23_transpose_y_0 = const()[name = tensor<string, []>("states_23_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 6, ?, 64]> value_states_5 = transpose(perm = value_states_5_perm_0, x = var_321)[name = tensor<string, []>("transpose_63")];
+            tensor<fp32, [1, 6, ?, 64]> states_23 = matmul(transpose_x = states_23_transpose_x_0, transpose_y = states_23_transpose_y_0, x = var_327, y = value_states_5)[name = tensor<string, []>("states_23")];
+            tensor<int32, [4]> var_331_perm_0 = const()[name = tensor<string, []>("op_331_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_333 = const()[name = tensor<string, []>("op_333"), val = tensor<int32, [3]>([1, -1, 384])];
+            tensor<fp32, [1, ?, 6, 64]> var_331 = transpose(perm = var_331_perm_0, x = states_23)[name = tensor<string, []>("transpose_60")];
+            tensor<fp32, [1, ?, 384]> input_37 = reshape(shape = var_333, x = var_331)[name = tensor<string, []>("input_37")];
+            tensor<fp32, [1, ?, 256]> input_39 = linear(bias = linear_3_bias_0, weight = decoder_block_1_layer_0_SelfAttention_o_weight, x = input_37)[name = tensor<string, []>("linear_14")];
+            tensor<fp32, [1, ?, 256]> hidden_states_31 = add(x = hidden_states_23, y = input_39)[name = tensor<string, []>("hidden_states_31")];
+            tensor<fp32, []> var_18_promoted_4 = const()[name = tensor<string, []>("op_18_promoted_4"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1, ?, 256]> var_349 = pow(x = hidden_states_31, y = var_18_promoted_4)[name = tensor<string, []>("op_349")];
+            tensor<int32, [1]> variance_9_axes_0 = const()[name = tensor<string, []>("variance_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> variance_9_keep_dims_0 = const()[name = tensor<string, []>("variance_9_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp32, [1, ?, 1]> variance_9 = reduce_mean(axes = variance_9_axes_0, keep_dims = variance_9_keep_dims_0, x = var_349)[name = tensor<string, []>("variance_9")];
+            tensor<fp32, []> var_352 = const()[name = tensor<string, []>("op_352"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
+            tensor<fp32, [1, ?, 1]> var_353 = add(x = variance_9, y = var_352)[name = tensor<string, []>("op_353")];
+            tensor<fp32, []> var_354_epsilon_0 = const()[name = tensor<string, []>("op_354_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp32, [1, ?, 1]> var_354 = rsqrt(epsilon = var_354_epsilon_0, x = var_353)[name = tensor<string, []>("op_354")];
+            tensor<fp32, [1, ?, 256]> hidden_states_35 = mul(x = hidden_states_31, y = var_354)[name = tensor<string, []>("hidden_states_35")];
+            tensor<fp32, [1, ?, 256]> hidden_states_37 = mul(x = decoder_block_1_layer_1_layer_norm_weight, y = hidden_states_35)[name = tensor<string, []>("hidden_states_37")];
+            tensor<fp32, [1, ?, 384]> states_25 = linear(bias = linear_0_bias_0, weight = decoder_block_1_layer_1_EncDecAttention_q_weight, x = hidden_states_37)[name = tensor<string, []>("linear_15")];
+            tensor<int32, [4]> var_367 = const()[name = tensor<string, []>("op_367"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_368 = reshape(shape = var_367, x = states_25)[name = tensor<string, []>("op_368")];
+            tensor<fp32, [1, ?, 384]> states_27 = linear(bias = linear_0_bias_0, weight = decoder_block_1_layer_1_EncDecAttention_k_weight, x = encoder_hidden_states)[name = tensor<string, []>("linear_16")];
+            tensor<int32, [4]> var_372 = const()[name = tensor<string, []>("op_372"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_373 = reshape(shape = var_372, x = states_27)[name = tensor<string, []>("op_373")];
+            tensor<fp32, [1, ?, 384]> states_29 = linear(bias = linear_0_bias_0, weight = decoder_block_1_layer_1_EncDecAttention_v_weight, x = encoder_hidden_states)[name = tensor<string, []>("linear_17")];
+            tensor<int32, [4]> var_377 = const()[name = tensor<string, []>("op_377"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_378 = reshape(shape = var_377, x = states_29)[name = tensor<string, []>("op_378")];
+            tensor<int32, [4]> value_states_7_perm_0 = const()[name = tensor<string, []>("value_states_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> scores_13_transpose_x_0 = const()[name = tensor<string, []>("scores_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> scores_13_transpose_y_0 = const()[name = tensor<string, []>("scores_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_30_perm_0 = const()[name = tensor<string, []>("transpose_30_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_31_perm_0 = const()[name = tensor<string, []>("transpose_31_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 6, 64, ?]> transpose_31 = transpose(perm = transpose_31_perm_0, x = var_373)[name = tensor<string, []>("transpose_57")];
+            tensor<fp32, [1, 6, ?, 64]> transpose_30 = transpose(perm = transpose_30_perm_0, x = var_368)[name = tensor<string, []>("transpose_58")];
+            tensor<fp32, [1, 6, ?, ?]> scores_13 = matmul(transpose_x = scores_13_transpose_x_0, transpose_y = scores_13_transpose_y_0, x = transpose_30, y = transpose_31)[name = tensor<string, []>("scores_13")];
+            tensor<fp32, [1, 6, ?, ?]> scores_15 = add(x = scores_13, y = position_bias)[name = tensor<string, []>("scores_15")];
+            tensor<fp32, [1, 6, ?, ?]> var_384 = softmax(axis = var_22, x = scores_15)[name = tensor<string, []>("op_384")];
+            tensor<bool, []> states_31_transpose_x_0 = const()[name = tensor<string, []>("states_31_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> states_31_transpose_y_0 = const()[name = tensor<string, []>("states_31_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 6, ?, 64]> value_states_7 = transpose(perm = value_states_7_perm_0, x = var_378)[name = tensor<string, []>("transpose_59")];
+            tensor<fp32, [1, 6, ?, 64]> states_31 = matmul(transpose_x = states_31_transpose_x_0, transpose_y = states_31_transpose_y_0, x = var_384, y = value_states_7)[name = tensor<string, []>("states_31")];
+            tensor<int32, [4]> var_388_perm_0 = const()[name = tensor<string, []>("op_388_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_390 = const()[name = tensor<string, []>("op_390"), val = tensor<int32, [3]>([1, -1, 384])];
+            tensor<fp32, [1, ?, 6, 64]> var_388 = transpose(perm = var_388_perm_0, x = states_31)[name = tensor<string, []>("transpose_56")];
+            tensor<fp32, [1, ?, 384]> input_45 = reshape(shape = var_390, x = var_388)[name = tensor<string, []>("input_45")];
+            tensor<fp32, [1, ?, 256]> input_47 = linear(bias = linear_3_bias_0, weight = decoder_block_1_layer_1_EncDecAttention_o_weight, x = input_45)[name = tensor<string, []>("linear_18")];
+            tensor<fp32, [1, ?, 256]> hidden_states_39 = add(x = hidden_states_31, y = input_47)[name = tensor<string, []>("hidden_states_39")];
+            tensor<fp32, []> var_18_promoted_5 = const()[name = tensor<string, []>("op_18_promoted_5"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1, ?, 256]> var_400 = pow(x = hidden_states_39, y = var_18_promoted_5)[name = tensor<string, []>("op_400")];
+            tensor<int32, [1]> variance_11_axes_0 = const()[name = tensor<string, []>("variance_11_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> variance_11_keep_dims_0 = const()[name = tensor<string, []>("variance_11_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp32, [1, ?, 1]> variance_11 = reduce_mean(axes = variance_11_axes_0, keep_dims = variance_11_keep_dims_0, x = var_400)[name = tensor<string, []>("variance_11")];
+            tensor<fp32, []> var_403 = const()[name = tensor<string, []>("op_403"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
+            tensor<fp32, [1, ?, 1]> var_404 = add(x = variance_11, y = var_403)[name = tensor<string, []>("op_404")];
+            tensor<fp32, []> var_405_epsilon_0 = const()[name = tensor<string, []>("op_405_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp32, [1, ?, 1]> var_405 = rsqrt(epsilon = var_405_epsilon_0, x = var_404)[name = tensor<string, []>("op_405")];
+            tensor<fp32, [1, ?, 256]> hidden_states_43 = mul(x = hidden_states_39, y = var_405)[name = tensor<string, []>("hidden_states_43")];
+            tensor<fp32, [1, ?, 256]> input_49 = mul(x = decoder_block_1_layer_2_layer_norm_weight, y = hidden_states_43)[name = tensor<string, []>("input_49")];
+            tensor<fp32, [1, ?, 1024]> input_51 = linear(bias = linear_8_bias_0, weight = decoder_block_1_layer_2_DenseReluDense_wi_0_weight, x = input_49)[name = tensor<string, []>("linear_19")];
+            tensor<string, []> hidden_gelu_3_mode_0 = const()[name = tensor<string, []>("hidden_gelu_3_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
+            tensor<fp32, [1, ?, 1024]> hidden_gelu_3 = gelu(mode = hidden_gelu_3_mode_0, x = input_51)[name = tensor<string, []>("hidden_gelu_3")];
+            tensor<fp32, [1, ?, 1024]> hidden_linear_3 = linear(bias = linear_8_bias_0, weight = decoder_block_1_layer_2_DenseReluDense_wi_1_weight, x = input_49)[name = tensor<string, []>("linear_20")];
+            tensor<fp32, [1, ?, 1024]> input_53 = mul(x = hidden_gelu_3, y = hidden_linear_3)[name = tensor<string, []>("input_53")];
+            tensor<fp32, [1, ?, 256]> input_57 = linear(bias = linear_3_bias_0, weight = decoder_block_1_layer_2_DenseReluDense_wo_weight, x = input_53)[name = tensor<string, []>("linear_21")];
+            tensor<fp32, [1, ?, 256]> hidden_states_45 = add(x = hidden_states_39, y = input_57)[name = tensor<string, []>("hidden_states_45")];
+            tensor<fp32, []> var_18_promoted_6 = const()[name = tensor<string, []>("op_18_promoted_6"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1, ?, 256]> var_446 = pow(x = hidden_states_45, y = var_18_promoted_6)[name = tensor<string, []>("op_446")];
+            tensor<int32, [1]> variance_13_axes_0 = const()[name = tensor<string, []>("variance_13_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> variance_13_keep_dims_0 = const()[name = tensor<string, []>("variance_13_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp32, [1, ?, 1]> variance_13 = reduce_mean(axes = variance_13_axes_0, keep_dims = variance_13_keep_dims_0, x = var_446)[name = tensor<string, []>("variance_13")];
+            tensor<fp32, []> var_449 = const()[name = tensor<string, []>("op_449"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
+            tensor<fp32, [1, ?, 1]> var_450 = add(x = variance_13, y = var_449)[name = tensor<string, []>("op_450")];
+            tensor<fp32, []> var_451_epsilon_0 = const()[name = tensor<string, []>("op_451_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp32, [1, ?, 1]> var_451 = rsqrt(epsilon = var_451_epsilon_0, x = var_450)[name = tensor<string, []>("op_451")];
+            tensor<fp32, [1, ?, 256]> hidden_states_49 = mul(x = hidden_states_45, y = var_451)[name = tensor<string, []>("hidden_states_49")];
+            tensor<fp32, [1, ?, 256]> hidden_states_51 = mul(x = decoder_block_2_layer_0_layer_norm_weight, y = hidden_states_49)[name = tensor<string, []>("hidden_states_51")];
+            tensor<fp32, [1, ?, 384]> states_33 = linear(bias = linear_0_bias_0, weight = decoder_block_2_layer_0_SelfAttention_q_weight, x = hidden_states_51)[name = tensor<string, []>("linear_22")];
+            tensor<int32, [4]> var_464 = const()[name = tensor<string, []>("op_464"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_465 = reshape(shape = var_464, x = states_33)[name = tensor<string, []>("op_465")];
+            tensor<fp32, [1, ?, 384]> states_35 = linear(bias = linear_0_bias_0, weight = decoder_block_2_layer_0_SelfAttention_k_weight, x = hidden_states_51)[name = tensor<string, []>("linear_23")];
+            tensor<int32, [4]> var_469 = const()[name = tensor<string, []>("op_469"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_470 = reshape(shape = var_469, x = states_35)[name = tensor<string, []>("op_470")];
+            tensor<fp32, [1, ?, 384]> states_37 = linear(bias = linear_0_bias_0, weight = decoder_block_2_layer_0_SelfAttention_v_weight, x = hidden_states_51)[name = tensor<string, []>("linear_24")];
+            tensor<int32, [4]> var_474 = const()[name = tensor<string, []>("op_474"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_475 = reshape(shape = var_474, x = states_37)[name = tensor<string, []>("op_475")];
+            tensor<int32, [4]> value_states_9_perm_0 = const()[name = tensor<string, []>("value_states_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> scores_17_transpose_x_0 = const()[name = tensor<string, []>("scores_17_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> scores_17_transpose_y_0 = const()[name = tensor<string, []>("scores_17_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_32_perm_0 = const()[name = tensor<string, []>("transpose_32_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_33_perm_0 = const()[name = tensor<string, []>("transpose_33_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 6, 64, ?]> transpose_33 = transpose(perm = transpose_33_perm_0, x = var_470)[name = tensor<string, []>("transpose_53")];
+            tensor<fp32, [1, 6, ?, 64]> transpose_32 = transpose(perm = transpose_32_perm_0, x = var_465)[name = tensor<string, []>("transpose_54")];
+            tensor<fp32, [1, 6, ?, ?]> scores_17 = matmul(transpose_x = scores_17_transpose_x_0, transpose_y = scores_17_transpose_y_0, x = transpose_32, y = transpose_33)[name = tensor<string, []>("scores_17")];
+            tensor<fp32, [1, 6, ?, ?]> scores_19 = add(x = scores_17, y = position_bias_3)[name = tensor<string, []>("scores_19")];
+            tensor<fp32, [1, 6, ?, ?]> var_481 = softmax(axis = var_22, x = scores_19)[name = tensor<string, []>("op_481")];
+            tensor<bool, []> states_39_transpose_x_0 = const()[name = tensor<string, []>("states_39_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> states_39_transpose_y_0 = const()[name = tensor<string, []>("states_39_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 6, ?, 64]> value_states_9 = transpose(perm = value_states_9_perm_0, x = var_475)[name = tensor<string, []>("transpose_55")];
+            tensor<fp32, [1, 6, ?, 64]> states_39 = matmul(transpose_x = states_39_transpose_x_0, transpose_y = states_39_transpose_y_0, x = var_481, y = value_states_9)[name = tensor<string, []>("states_39")];
+            tensor<int32, [4]> var_485_perm_0 = const()[name = tensor<string, []>("op_485_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_487 = const()[name = tensor<string, []>("op_487"), val = tensor<int32, [3]>([1, -1, 384])];
+            tensor<fp32, [1, ?, 6, 64]> var_485 = transpose(perm = var_485_perm_0, x = states_39)[name = tensor<string, []>("transpose_52")];
+            tensor<fp32, [1, ?, 384]> input_63 = reshape(shape = var_487, x = var_485)[name = tensor<string, []>("input_63")];
+            tensor<fp32, [1, ?, 256]> input_65 = linear(bias = linear_3_bias_0, weight = decoder_block_2_layer_0_SelfAttention_o_weight, x = input_63)[name = tensor<string, []>("linear_25")];
+            tensor<fp32, [1, ?, 256]> hidden_states_53 = add(x = hidden_states_45, y = input_65)[name = tensor<string, []>("hidden_states_53")];
+            tensor<fp32, []> var_18_promoted_7 = const()[name = tensor<string, []>("op_18_promoted_7"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1, ?, 256]> var_503 = pow(x = hidden_states_53, y = var_18_promoted_7)[name = tensor<string, []>("op_503")];
+            tensor<int32, [1]> variance_15_axes_0 = const()[name = tensor<string, []>("variance_15_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> variance_15_keep_dims_0 = const()[name = tensor<string, []>("variance_15_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp32, [1, ?, 1]> variance_15 = reduce_mean(axes = variance_15_axes_0, keep_dims = variance_15_keep_dims_0, x = var_503)[name = tensor<string, []>("variance_15")];
+            tensor<fp32, []> var_506 = const()[name = tensor<string, []>("op_506"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
+            tensor<fp32, [1, ?, 1]> var_507 = add(x = variance_15, y = var_506)[name = tensor<string, []>("op_507")];
+            tensor<fp32, []> var_508_epsilon_0 = const()[name = tensor<string, []>("op_508_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp32, [1, ?, 1]> var_508 = rsqrt(epsilon = var_508_epsilon_0, x = var_507)[name = tensor<string, []>("op_508")];
+            tensor<fp32, [1, ?, 256]> hidden_states_57 = mul(x = hidden_states_53, y = var_508)[name = tensor<string, []>("hidden_states_57")];
+            tensor<fp32, [1, ?, 256]> hidden_states_59 = mul(x = decoder_block_2_layer_1_layer_norm_weight, y = hidden_states_57)[name = tensor<string, []>("hidden_states_59")];
+            tensor<fp32, [1, ?, 384]> states_41 = linear(bias = linear_0_bias_0, weight = decoder_block_2_layer_1_EncDecAttention_q_weight, x = hidden_states_59)[name = tensor<string, []>("linear_26")];
+            tensor<int32, [4]> var_521 = const()[name = tensor<string, []>("op_521"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_522 = reshape(shape = var_521, x = states_41)[name = tensor<string, []>("op_522")];
+            tensor<fp32, [1, ?, 384]> states_43 = linear(bias = linear_0_bias_0, weight = decoder_block_2_layer_1_EncDecAttention_k_weight, x = encoder_hidden_states)[name = tensor<string, []>("linear_27")];
+            tensor<int32, [4]> var_526 = const()[name = tensor<string, []>("op_526"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_527 = reshape(shape = var_526, x = states_43)[name = tensor<string, []>("op_527")];
+            tensor<fp32, [1, ?, 384]> states_45 = linear(bias = linear_0_bias_0, weight = decoder_block_2_layer_1_EncDecAttention_v_weight, x = encoder_hidden_states)[name = tensor<string, []>("linear_28")];
+            tensor<int32, [4]> var_531 = const()[name = tensor<string, []>("op_531"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_532 = reshape(shape = var_531, x = states_45)[name = tensor<string, []>("op_532")];
+            tensor<int32, [4]> value_states_11_perm_0 = const()[name = tensor<string, []>("value_states_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> scores_21_transpose_x_0 = const()[name = tensor<string, []>("scores_21_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> scores_21_transpose_y_0 = const()[name = tensor<string, []>("scores_21_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_34_perm_0 = const()[name = tensor<string, []>("transpose_34_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_35_perm_0 = const()[name = tensor<string, []>("transpose_35_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 6, 64, ?]> transpose_35 = transpose(perm = transpose_35_perm_0, x = var_527)[name = tensor<string, []>("transpose_49")];
+            tensor<fp32, [1, 6, ?, 64]> transpose_34 = transpose(perm = transpose_34_perm_0, x = var_522)[name = tensor<string, []>("transpose_50")];
+            tensor<fp32, [1, 6, ?, ?]> scores_21 = matmul(transpose_x = scores_21_transpose_x_0, transpose_y = scores_21_transpose_y_0, x = transpose_34, y = transpose_35)[name = tensor<string, []>("scores_21")];
+            tensor<fp32, [1, 6, ?, ?]> scores_23 = add(x = scores_21, y = position_bias)[name = tensor<string, []>("scores_23")];
+            tensor<fp32, [1, 6, ?, ?]> var_538 = softmax(axis = var_22, x = scores_23)[name = tensor<string, []>("op_538")];
+            tensor<bool, []> states_47_transpose_x_0 = const()[name = tensor<string, []>("states_47_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> states_47_transpose_y_0 = const()[name = tensor<string, []>("states_47_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 6, ?, 64]> value_states_11 = transpose(perm = value_states_11_perm_0, x = var_532)[name = tensor<string, []>("transpose_51")];
+            tensor<fp32, [1, 6, ?, 64]> states_47 = matmul(transpose_x = states_47_transpose_x_0, transpose_y = states_47_transpose_y_0, x = var_538, y = value_states_11)[name = tensor<string, []>("states_47")];
+            tensor<int32, [4]> var_542_perm_0 = const()[name = tensor<string, []>("op_542_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_544 = const()[name = tensor<string, []>("op_544"), val = tensor<int32, [3]>([1, -1, 384])];
+            tensor<fp32, [1, ?, 6, 64]> var_542 = transpose(perm = var_542_perm_0, x = states_47)[name = tensor<string, []>("transpose_48")];
+            tensor<fp32, [1, ?, 384]> input_71 = reshape(shape = var_544, x = var_542)[name = tensor<string, []>("input_71")];
+            tensor<fp32, [1, ?, 256]> input_73 = linear(bias = linear_3_bias_0, weight = decoder_block_2_layer_1_EncDecAttention_o_weight, x = input_71)[name = tensor<string, []>("linear_29")];
+            tensor<fp32, [1, ?, 256]> hidden_states_61 = add(x = hidden_states_53, y = input_73)[name = tensor<string, []>("hidden_states_61")];
+            tensor<fp32, []> var_18_promoted_8 = const()[name = tensor<string, []>("op_18_promoted_8"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1, ?, 256]> var_554 = pow(x = hidden_states_61, y = var_18_promoted_8)[name = tensor<string, []>("op_554")];
+            tensor<int32, [1]> variance_17_axes_0 = const()[name = tensor<string, []>("variance_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> variance_17_keep_dims_0 = const()[name = tensor<string, []>("variance_17_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp32, [1, ?, 1]> variance_17 = reduce_mean(axes = variance_17_axes_0, keep_dims = variance_17_keep_dims_0, x = var_554)[name = tensor<string, []>("variance_17")];
+            tensor<fp32, []> var_557 = const()[name = tensor<string, []>("op_557"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
+            tensor<fp32, [1, ?, 1]> var_558 = add(x = variance_17, y = var_557)[name = tensor<string, []>("op_558")];
+            tensor<fp32, []> var_559_epsilon_0 = const()[name = tensor<string, []>("op_559_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp32, [1, ?, 1]> var_559 = rsqrt(epsilon = var_559_epsilon_0, x = var_558)[name = tensor<string, []>("op_559")];
+            tensor<fp32, [1, ?, 256]> hidden_states_65 = mul(x = hidden_states_61, y = var_559)[name = tensor<string, []>("hidden_states_65")];
+            tensor<fp32, [1, ?, 256]> input_75 = mul(x = decoder_block_2_layer_2_layer_norm_weight, y = hidden_states_65)[name = tensor<string, []>("input_75")];
+            tensor<fp32, [1, ?, 1024]> input_77 = linear(bias = linear_8_bias_0, weight = decoder_block_2_layer_2_DenseReluDense_wi_0_weight, x = input_75)[name = tensor<string, []>("linear_30")];
+            tensor<string, []> hidden_gelu_5_mode_0 = const()[name = tensor<string, []>("hidden_gelu_5_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
+            tensor<fp32, [1, ?, 1024]> hidden_gelu_5 = gelu(mode = hidden_gelu_5_mode_0, x = input_77)[name = tensor<string, []>("hidden_gelu_5")];
+            tensor<fp32, [1, ?, 1024]> hidden_linear_5 = linear(bias = linear_8_bias_0, weight = decoder_block_2_layer_2_DenseReluDense_wi_1_weight, x = input_75)[name = tensor<string, []>("linear_31")];
+            tensor<fp32, [1, ?, 1024]> input_79 = mul(x = hidden_gelu_5, y = hidden_linear_5)[name = tensor<string, []>("input_79")];
+            tensor<fp32, [1, ?, 256]> input_83 = linear(bias = linear_3_bias_0, weight = decoder_block_2_layer_2_DenseReluDense_wo_weight, x = input_79)[name = tensor<string, []>("linear_32")];
+            tensor<fp32, [1, ?, 256]> hidden_states_67 = add(x = hidden_states_61, y = input_83)[name = tensor<string, []>("hidden_states_67")];
+            tensor<fp32, []> var_18_promoted_9 = const()[name = tensor<string, []>("op_18_promoted_9"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1, ?, 256]> var_600 = pow(x = hidden_states_67, y = var_18_promoted_9)[name = tensor<string, []>("op_600")];
+            tensor<int32, [1]> variance_19_axes_0 = const()[name = tensor<string, []>("variance_19_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> variance_19_keep_dims_0 = const()[name = tensor<string, []>("variance_19_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp32, [1, ?, 1]> variance_19 = reduce_mean(axes = variance_19_axes_0, keep_dims = variance_19_keep_dims_0, x = var_600)[name = tensor<string, []>("variance_19")];
+            tensor<fp32, []> var_603 = const()[name = tensor<string, []>("op_603"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
+            tensor<fp32, [1, ?, 1]> var_604 = add(x = variance_19, y = var_603)[name = tensor<string, []>("op_604")];
+            tensor<fp32, []> var_605_epsilon_0 = const()[name = tensor<string, []>("op_605_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp32, [1, ?, 1]> var_605 = rsqrt(epsilon = var_605_epsilon_0, x = var_604)[name = tensor<string, []>("op_605")];
+            tensor<fp32, [1, ?, 256]> hidden_states_71 = mul(x = hidden_states_67, y = var_605)[name = tensor<string, []>("hidden_states_71")];
+            tensor<fp32, [1, ?, 256]> hidden_states_73 = mul(x = decoder_block_3_layer_0_layer_norm_weight, y = hidden_states_71)[name = tensor<string, []>("hidden_states_73")];
+            tensor<fp32, [1, ?, 384]> states_49 = linear(bias = linear_0_bias_0, weight = decoder_block_3_layer_0_SelfAttention_q_weight, x = hidden_states_73)[name = tensor<string, []>("linear_33")];
+            tensor<int32, [4]> var_618 = const()[name = tensor<string, []>("op_618"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_619 = reshape(shape = var_618, x = states_49)[name = tensor<string, []>("op_619")];
+            tensor<fp32, [1, ?, 384]> states_51 = linear(bias = linear_0_bias_0, weight = decoder_block_3_layer_0_SelfAttention_k_weight, x = hidden_states_73)[name = tensor<string, []>("linear_34")];
+            tensor<int32, [4]> var_623 = const()[name = tensor<string, []>("op_623"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_624 = reshape(shape = var_623, x = states_51)[name = tensor<string, []>("op_624")];
+            tensor<fp32, [1, ?, 384]> states_53 = linear(bias = linear_0_bias_0, weight = decoder_block_3_layer_0_SelfAttention_v_weight, x = hidden_states_73)[name = tensor<string, []>("linear_35")];
+            tensor<int32, [4]> var_628 = const()[name = tensor<string, []>("op_628"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_629 = reshape(shape = var_628, x = states_53)[name = tensor<string, []>("op_629")];
+            tensor<int32, [4]> value_states_13_perm_0 = const()[name = tensor<string, []>("value_states_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> scores_25_transpose_x_0 = const()[name = tensor<string, []>("scores_25_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> scores_25_transpose_y_0 = const()[name = tensor<string, []>("scores_25_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_36_perm_0 = const()[name = tensor<string, []>("transpose_36_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_37_perm_0 = const()[name = tensor<string, []>("transpose_37_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 6, 64, ?]> transpose_37 = transpose(perm = transpose_37_perm_0, x = var_624)[name = tensor<string, []>("transpose_45")];
+            tensor<fp32, [1, 6, ?, 64]> transpose_36 = transpose(perm = transpose_36_perm_0, x = var_619)[name = tensor<string, []>("transpose_46")];
+            tensor<fp32, [1, 6, ?, ?]> scores_25 = matmul(transpose_x = scores_25_transpose_x_0, transpose_y = scores_25_transpose_y_0, x = transpose_36, y = transpose_37)[name = tensor<string, []>("scores_25")];
+            tensor<fp32, [1, 6, ?, ?]> scores_27 = add(x = scores_25, y = position_bias_3)[name = tensor<string, []>("scores_27")];
+            tensor<fp32, [1, 6, ?, ?]> var_635 = softmax(axis = var_22, x = scores_27)[name = tensor<string, []>("op_635")];
+            tensor<bool, []> states_55_transpose_x_0 = const()[name = tensor<string, []>("states_55_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> states_55_transpose_y_0 = const()[name = tensor<string, []>("states_55_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 6, ?, 64]> value_states_13 = transpose(perm = value_states_13_perm_0, x = var_629)[name = tensor<string, []>("transpose_47")];
+            tensor<fp32, [1, 6, ?, 64]> states_55 = matmul(transpose_x = states_55_transpose_x_0, transpose_y = states_55_transpose_y_0, x = var_635, y = value_states_13)[name = tensor<string, []>("states_55")];
+            tensor<int32, [4]> var_639_perm_0 = const()[name = tensor<string, []>("op_639_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_641 = const()[name = tensor<string, []>("op_641"), val = tensor<int32, [3]>([1, -1, 384])];
+            tensor<fp32, [1, ?, 6, 64]> var_639 = transpose(perm = var_639_perm_0, x = states_55)[name = tensor<string, []>("transpose_44")];
+            tensor<fp32, [1, ?, 384]> input_89 = reshape(shape = var_641, x = var_639)[name = tensor<string, []>("input_89")];
+            tensor<fp32, [1, ?, 256]> input_91 = linear(bias = linear_3_bias_0, weight = decoder_block_3_layer_0_SelfAttention_o_weight, x = input_89)[name = tensor<string, []>("linear_36")];
+            tensor<fp32, [1, ?, 256]> hidden_states_75 = add(x = hidden_states_67, y = input_91)[name = tensor<string, []>("hidden_states_75")];
+            tensor<fp32, []> var_18_promoted_10 = const()[name = tensor<string, []>("op_18_promoted_10"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1, ?, 256]> var_657 = pow(x = hidden_states_75, y = var_18_promoted_10)[name = tensor<string, []>("op_657")];
+            tensor<int32, [1]> variance_21_axes_0 = const()[name = tensor<string, []>("variance_21_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> variance_21_keep_dims_0 = const()[name = tensor<string, []>("variance_21_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp32, [1, ?, 1]> variance_21 = reduce_mean(axes = variance_21_axes_0, keep_dims = variance_21_keep_dims_0, x = var_657)[name = tensor<string, []>("variance_21")];
+            tensor<fp32, []> var_660 = const()[name = tensor<string, []>("op_660"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
+            tensor<fp32, [1, ?, 1]> var_661 = add(x = variance_21, y = var_660)[name = tensor<string, []>("op_661")];
+            tensor<fp32, []> var_662_epsilon_0 = const()[name = tensor<string, []>("op_662_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp32, [1, ?, 1]> var_662 = rsqrt(epsilon = var_662_epsilon_0, x = var_661)[name = tensor<string, []>("op_662")];
+            tensor<fp32, [1, ?, 256]> hidden_states_79 = mul(x = hidden_states_75, y = var_662)[name = tensor<string, []>("hidden_states_79")];
+            tensor<fp32, [1, ?, 256]> hidden_states_81 = mul(x = decoder_block_3_layer_1_layer_norm_weight, y = hidden_states_79)[name = tensor<string, []>("hidden_states_81")];
+            tensor<fp32, [1, ?, 384]> states_57 = linear(bias = linear_0_bias_0, weight = decoder_block_3_layer_1_EncDecAttention_q_weight, x = hidden_states_81)[name = tensor<string, []>("linear_37")];
+            tensor<int32, [4]> var_675 = const()[name = tensor<string, []>("op_675"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_676 = reshape(shape = var_675, x = states_57)[name = tensor<string, []>("op_676")];
+            tensor<fp32, [1, ?, 384]> states_59 = linear(bias = linear_0_bias_0, weight = decoder_block_3_layer_1_EncDecAttention_k_weight, x = encoder_hidden_states)[name = tensor<string, []>("linear_38")];
+            tensor<int32, [4]> var_680 = const()[name = tensor<string, []>("op_680"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_681 = reshape(shape = var_680, x = states_59)[name = tensor<string, []>("op_681")];
+            tensor<fp32, [1, ?, 384]> states_61 = linear(bias = linear_0_bias_0, weight = decoder_block_3_layer_1_EncDecAttention_v_weight, x = encoder_hidden_states)[name = tensor<string, []>("linear_39")];
+            tensor<int32, [4]> var_685 = const()[name = tensor<string, []>("op_685"), val = tensor<int32, [4]>([1, -1, 6, 64])];
+            tensor<fp32, [1, ?, 6, 64]> var_686 = reshape(shape = var_685, x = states_61)[name = tensor<string, []>("op_686")];
+            tensor<int32, [4]> value_states_perm_0 = const()[name = tensor<string, []>("value_states_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<bool, []> scores_29_transpose_x_0 = const()[name = tensor<string, []>("scores_29_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> scores_29_transpose_y_0 = const()[name = tensor<string, []>("scores_29_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_38_perm_0 = const()[name = tensor<string, []>("transpose_38_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_39_perm_0 = const()[name = tensor<string, []>("transpose_39_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 6, 64, ?]> transpose_39 = transpose(perm = transpose_39_perm_0, x = var_681)[name = tensor<string, []>("transpose_41")];
+            tensor<fp32, [1, 6, ?, 64]> transpose_38 = transpose(perm = transpose_38_perm_0, x = var_676)[name = tensor<string, []>("transpose_42")];
+            tensor<fp32, [1, 6, ?, ?]> scores_29 = matmul(transpose_x = scores_29_transpose_x_0, transpose_y = scores_29_transpose_y_0, x = transpose_38, y = transpose_39)[name = tensor<string, []>("scores_29")];
+            tensor<fp32, [1, 6, ?, ?]> scores = add(x = scores_29, y = position_bias)[name = tensor<string, []>("scores")];
+            tensor<fp32, [1, 6, ?, ?]> var_692 = softmax(axis = var_22, x = scores)[name = tensor<string, []>("op_692")];
+            tensor<bool, []> states_transpose_x_0 = const()[name = tensor<string, []>("states_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> states_transpose_y_0 = const()[name = tensor<string, []>("states_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 6, ?, 64]> value_states = transpose(perm = value_states_perm_0, x = var_686)[name = tensor<string, []>("transpose_43")];
+            tensor<fp32, [1, 6, ?, 64]> states = matmul(transpose_x = states_transpose_x_0, transpose_y = states_transpose_y_0, x = var_692, y = value_states)[name = tensor<string, []>("states")];
+            tensor<int32, [4]> var_696_perm_0 = const()[name = tensor<string, []>("op_696_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_698 = const()[name = tensor<string, []>("op_698"), val = tensor<int32, [3]>([1, -1, 384])];
+            tensor<fp32, [1, ?, 6, 64]> var_696 = transpose(perm = var_696_perm_0, x = states)[name = tensor<string, []>("transpose_40")];
+            tensor<fp32, [1, ?, 384]> input_97 = reshape(shape = var_698, x = var_696)[name = tensor<string, []>("input_97")];
+            tensor<fp32, [1, ?, 256]> input_99 = linear(bias = linear_3_bias_0, weight = decoder_block_3_layer_1_EncDecAttention_o_weight, x = input_97)[name = tensor<string, []>("linear_40")];
+            tensor<fp32, [1, ?, 256]> hidden_states_83 = add(x = hidden_states_75, y = input_99)[name = tensor<string, []>("hidden_states_83")];
+            tensor<fp32, []> var_18_promoted_11 = const()[name = tensor<string, []>("op_18_promoted_11"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1, ?, 256]> var_708 = pow(x = hidden_states_83, y = var_18_promoted_11)[name = tensor<string, []>("op_708")];
+            tensor<int32, [1]> variance_23_axes_0 = const()[name = tensor<string, []>("variance_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> variance_23_keep_dims_0 = const()[name = tensor<string, []>("variance_23_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp32, [1, ?, 1]> variance_23 = reduce_mean(axes = variance_23_axes_0, keep_dims = variance_23_keep_dims_0, x = var_708)[name = tensor<string, []>("variance_23")];
+            tensor<fp32, []> var_711 = const()[name = tensor<string, []>("op_711"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
+            tensor<fp32, [1, ?, 1]> var_712 = add(x = variance_23, y = var_711)[name = tensor<string, []>("op_712")];
+            tensor<fp32, []> var_713_epsilon_0 = const()[name = tensor<string, []>("op_713_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp32, [1, ?, 1]> var_713 = rsqrt(epsilon = var_713_epsilon_0, x = var_712)[name = tensor<string, []>("op_713")];
+            tensor<fp32, [1, ?, 256]> hidden_states_87 = mul(x = hidden_states_83, y = var_713)[name = tensor<string, []>("hidden_states_87")];
+            tensor<fp32, [1, ?, 256]> input_101 = mul(x = decoder_block_3_layer_2_layer_norm_weight, y = hidden_states_87)[name = tensor<string, []>("input_101")];
+            tensor<fp32, [1, ?, 1024]> input_103 = linear(bias = linear_8_bias_0, weight = decoder_block_3_layer_2_DenseReluDense_wi_0_weight, x = input_101)[name = tensor<string, []>("linear_41")];
+            tensor<string, []> hidden_gelu_mode_0 = const()[name = tensor<string, []>("hidden_gelu_mode_0"), val = tensor<string, []>("TANH_APPROXIMATION")];
+            tensor<fp32, [1, ?, 1024]> hidden_gelu = gelu(mode = hidden_gelu_mode_0, x = input_103)[name = tensor<string, []>("hidden_gelu")];
+            tensor<fp32, [1, ?, 1024]> hidden_linear = linear(bias = linear_8_bias_0, weight = decoder_block_3_layer_2_DenseReluDense_wi_1_weight, x = input_101)[name = tensor<string, []>("linear_42")];
+            tensor<fp32, [1, ?, 1024]> input_105 = mul(x = hidden_gelu, y = hidden_linear)[name = tensor<string, []>("input_105")];
+            tensor<fp32, [1, ?, 256]> input_109 = linear(bias = linear_3_bias_0, weight = decoder_block_3_layer_2_DenseReluDense_wo_weight, x = input_105)[name = tensor<string, []>("linear_43")];
+            tensor<fp32, [1, ?, 256]> hidden_states_89 = add(x = hidden_states_83, y = input_109)[name = tensor<string, []>("hidden_states_89")];
+            tensor<fp32, []> var_18_promoted_12 = const()[name = tensor<string, []>("op_18_promoted_12"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1, ?, 256]> var_746 = pow(x = hidden_states_89, y = var_18_promoted_12)[name = tensor<string, []>("op_746")];
+            tensor<int32, [1]> variance_axes_0 = const()[name = tensor<string, []>("variance_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> variance_keep_dims_0 = const()[name = tensor<string, []>("variance_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp32, [1, ?, 1]> variance = reduce_mean(axes = variance_axes_0, keep_dims = variance_keep_dims_0, x = var_746)[name = tensor<string, []>("variance")];
+            tensor<fp32, []> var_749 = const()[name = tensor<string, []>("op_749"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
+            tensor<fp32, [1, ?, 1]> var_750 = add(x = variance, y = var_749)[name = tensor<string, []>("op_750")];
+            tensor<fp32, []> var_751_epsilon_0 = const()[name = tensor<string, []>("op_751_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
+            tensor<fp32, [1, ?, 1]> var_751 = rsqrt(epsilon = var_751_epsilon_0, x = var_750)[name = tensor<string, []>("op_751")];
+            tensor<fp32, [1, ?, 256]> hidden_states = mul(x = hidden_states_89, y = var_751)[name = tensor<string, []>("hidden_states")];
+            tensor<fp32, [1, ?, 256]> input_111 = mul(x = decoder_final_layer_norm_weight, y = hidden_states)[name = tensor<string, []>("input_111")];
+            tensor<fp32, [1, ?, 384]> logits = linear(bias = linear_0_bias_0, weight = lm_head_weight, x = input_111)[name = tensor<string, []>("linear_44")];
+        } -> (logits);
+}

MultilingualG2PDecoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:005b5ffbccabe4fd4074204d16af745b5d8f458a687b0c07e80060f7fe0a08e4
+size 25977088

MultilingualG2PEncoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee8fd1ea46ec79fd7816493973deaa583cd867ce0c7d2dd55124d2368db094bd
+size 243

MultilingualG2PEncoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed8bf165b8d5939f995b29490872ee581c9834d5b22d3686fc5c7b300680b058
+size 439

MultilingualG2PEncoder.mlmodelc/metadata.json

@@ -0,0 +1,99 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float32",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "last_hidden_state",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Range1d" : 1,
+      "Ios17.reshape" : 48,
+      "Ios16.reduceMean" : 25,
+      "Ios16.softmax" : 12,
+      "Ios17.matmul" : 24,
+      "Ios17.transpose" : 49,
+      "Select" : 3,
+      "Ios17.expandDims" : 5,
+      "Ios17.add" : 66,
+      "Ios16.fillLike" : 1,
+      "Shape" : 1,
+      "Ios17.gather" : 3,
+      "Ios17.log" : 1,
+      "Ios17.less" : 1,
+      "Ios17.sub" : 2,
+      "Ios17.pow" : 25,
+      "Ios17.cast" : 4,
+      "Ios16.gelu" : 12,
+      "Ios17.linear" : 84,
+      "Ios17.abs" : 1,
+      "Ios17.greaterEqual" : 2,
+      "Ios17.minimum" : 1,
+      "Ios17.rsqrt" : 25,
+      "Ios17.greater" : 1,
+      "Ios17.mul" : 67
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-03-13",
+      "com.github.apple.coremltools.source" : "torch==2.7.0",
+      "com.github.apple.coremltools.version" : "9.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "inputSchema" : [
+      {
+        "dataType" : "Int32",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 1...64",
+        "shapeRange" : "[[1, 1], [1, 64]]",
+        "formattedType" : "MultiArray (Int32 1 × 16)",
+        "type" : "MultiArray",
+        "shape" : "[1, 16]",
+        "name" : "input_ids",
+        "shortDescription" : ""
+      },
+      {
+        "dataType" : "Int32",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 1...64",
+        "shapeRange" : "[[1, 1], [1, 64]]",
+        "formattedType" : "MultiArray (Int32 1 × 16)",
+        "type" : "MultiArray",
+        "shape" : "[1, 16]",
+        "name" : "attention_mask",
+        "shortDescription" : ""
+      }
+    ],
+    "generatedClassName" : "G2PEncoder",
+    "method" : "predict"
+  }
+]

MultilingualG2PEncoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11b225e93b6b1fe6fb85c9d4317724e1a033d6123fdc01dc2ad85f3c1f168eb3
+size 57056704