Hugging Face's logo

FluidInference
/
pocket-tts-coreml

Text-to-Speech
Core ML
English
tts
apple
on-device
Model card Files
xet
 Community
1
alexwengg commited on
Commit
b3b6695
·
verified ·
1 Parent(s): a726c96

Upload 8 files

Browse files
Files changed (8) hide show
  1. mimi_encoder.mlmodelc/analytics/coremldata.bin +3 -0
  2. mimi_encoder.mlmodelc/coremldata.bin +3 -0
  3. mimi_encoder.mlmodelc/metadata.json +95 -0
  4. mimi_encoder.mlmodelc/model.mil +585 -0
  5. mimi_encoder.mlmodelc/weights/weight.bin +3 -0
  6. mimi_encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel +3 -0
  7. mimi_encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin +3 -0
  8. mimi_encoder.mlpackage/Manifest.json +18 -0
mimi_encoder.mlmodelc/analytics/coremldata.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2a988bbb13a4def26551000104951060686984c549f8bff4a56a38094d6d2829
3
+ size 243
mimi_encoder.mlmodelc/coremldata.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:85d8ff9e2879d15c60fe0426147373c628eda0c517b6b7f85c6e35292f6a0880
3
+ size 396
mimi_encoder.mlmodelc/metadata.json ADDED
@@ -0,0 +1,95 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "metadataOutputVersion" : "3.0",
4
+ "storagePrecision" : "Float32",
5
+ "outputSchema" : [
6
+ {
7
+ "hasShapeFlexibility" : "0",
8
+ "isOptional" : "0",
9
+ "dataType" : "Float32",
10
+ "formattedType" : "MultiArray (Float32)",
11
+ "shortDescription" : "",
12
+ "shape" : "[]",
13
+ "name" : "conditioning",
14
+ "type" : "MultiArray"
15
+ }
16
+ ],
17
+ "modelParameters" : [
18
+
19
+ ],
20
+ "specificationVersion" : 8,
21
+ "mlProgramOperationTypeHistogram" : {
22
+ "Tile" : 3,
23
+ "Ios17.layerNorm" : 4,
24
+ "Ios17.gather" : 15,
25
+ "Stack" : 4,
26
+ "Ios17.greaterEqual" : 4,
27
+ "Ios17.sub" : 8,
28
+ "Ios17.elu" : 10,
29
+ "Ios17.equal" : 1,
30
+ "Pad" : 8,
31
+ "Fill" : 1,
32
+ "Shape" : 12,
33
+ "Ios17.sliceByIndex" : 9,
34
+ "Ios17.concat" : 8,
35
+ "Range1d" : 6,
36
+ "Ios17.expandDims" : 6,
37
+ "Ios17.less" : 2,
38
+ "Ios17.mul" : 26,
39
+ "Ios17.transpose" : 15,
40
+ "Ios17.matmul" : 4,
41
+ "Ios16.gelu" : 2,
42
+ "Ios17.conv" : 12,
43
+ "Ios17.reshape" : 17,
44
+ "Split" : 2,
45
+ "Ios17.squeeze" : 6,
46
+ "Ios17.cast" : 4,
47
+ "Ios17.sin" : 2,
48
+ "Ios17.linear" : 9,
49
+ "Ios17.realDiv" : 3,
50
+ "Select" : 3,
51
+ "Ios16.softmax" : 2,
52
+ "Ios17.add" : 16,
53
+ "Ios17.cos" : 2,
54
+ "Ios17.logicalAnd" : 4
55
+ },
56
+ "computePrecision" : "Mixed (Float32, Int32)",
57
+ "isUpdatable" : "0",
58
+ "stateSchema" : [
59
+
60
+ ],
61
+ "availability" : {
62
+ "macOS" : "14.0",
63
+ "tvOS" : "17.0",
64
+ "visionOS" : "1.0",
65
+ "watchOS" : "10.0",
66
+ "iOS" : "17.0",
67
+ "macCatalyst" : "17.0"
68
+ },
69
+ "modelType" : {
70
+ "name" : "MLModelType_mlProgram"
71
+ },
72
+ "userDefinedMetadata" : {
73
+ "com.github.apple.coremltools.conversion_date" : "2026-02-03",
74
+ "com.github.apple.coremltools.source" : "torch==2.9.1",
75
+ "com.github.apple.coremltools.version" : "9.0",
76
+ "com.github.apple.coremltools.source_dialect" : "TorchScript"
77
+ },
78
+ "inputSchema" : [
79
+ {
80
+ "dataType" : "Float32",
81
+ "hasShapeFlexibility" : "1",
82
+ "isOptional" : "0",
83
+ "shapeFlexibility" : "1 × 1 × 1920...1440000",
84
+ "shapeRange" : "[[1, 1], [1, 1], [1920, 1440000]]",
85
+ "formattedType" : "MultiArray (Float32 1 × 1 × 1920)",
86
+ "type" : "MultiArray",
87
+ "shape" : "[1, 1, 1920]",
88
+ "name" : "audio",
89
+ "shortDescription" : ""
90
+ }
91
+ ],
92
+ "generatedClassName" : "mimi_encoder",
93
+ "method" : "predict"
94
+ }
95
+ ]
mimi_encoder.mlmodelc/model.mil ADDED
@@ -0,0 +1,585 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ program(1.0)
2
+ [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
3
+ {
4
+ func main<ios17>(tensor<fp32, [1, 1, ?]> audio) [FlexibleShapeInformation = tuple<tuple<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>, tuple<tensor<string, []>, dict<tensor<string, []>, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"audio", [1, 1, 1920]}}), ("RangeDims", {{"audio", [[1, 1], [1, 1], [1920, 1440000]]}})))] {
5
+ tensor<fp32, [1024, 512]> speaker_proj_weight = const()[name = tensor<string, []>("speaker_proj_weight"), val = tensor<fp32, [1024, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
6
+ tensor<fp32, [64]> encoder_model_0_conv_bias = const()[name = tensor<string, []>("encoder_model_0_conv_bias"), val = tensor<fp32, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2097280)))];
7
+ tensor<fp32, [64, 1, 7]> encoder_model_0_conv_weight = const()[name = tensor<string, []>("encoder_model_0_conv_weight"), val = tensor<fp32, [64, 1, 7]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2097600)))];
8
+ tensor<fp32, [32]> encoder_model_1_block_1_conv_bias = const()[name = tensor<string, []>("encoder_model_1_block_1_conv_bias"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2099456)))];
9
+ tensor<fp32, [32, 64, 3]> encoder_model_1_block_1_conv_weight = const()[name = tensor<string, []>("encoder_model_1_block_1_conv_weight"), val = tensor<fp32, [32, 64, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2099648)))];
10
+ tensor<fp32, [64]> encoder_model_1_block_3_conv_bias = const()[name = tensor<string, []>("encoder_model_1_block_3_conv_bias"), val = tensor<fp32, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2124288)))];
11
+ tensor<fp32, [64, 32, 1]> encoder_model_1_block_3_conv_weight = const()[name = tensor<string, []>("encoder_model_1_block_3_conv_weight"), val = tensor<fp32, [64, 32, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2124608)))];
12
+ tensor<fp32, [128]> encoder_model_3_conv_bias = const()[name = tensor<string, []>("encoder_model_3_conv_bias"), val = tensor<fp32, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2132864)))];
13
+ tensor<fp32, [128, 64, 8]> encoder_model_3_conv_weight = const()[name = tensor<string, []>("encoder_model_3_conv_weight"), val = tensor<fp32, [128, 64, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2133440)))];
14
+ tensor<fp32, [64]> encoder_model_4_block_1_conv_bias = const()[name = tensor<string, []>("encoder_model_4_block_1_conv_bias"), val = tensor<fp32, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2395648)))];
15
+ tensor<fp32, [64, 128, 3]> encoder_model_4_block_1_conv_weight = const()[name = tensor<string, []>("encoder_model_4_block_1_conv_weight"), val = tensor<fp32, [64, 128, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2395968)))];
16
+ tensor<fp32, [128]> encoder_model_4_block_3_conv_bias = const()[name = tensor<string, []>("encoder_model_4_block_3_conv_bias"), val = tensor<fp32, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2494336)))];
17
+ tensor<fp32, [128, 64, 1]> encoder_model_4_block_3_conv_weight = const()[name = tensor<string, []>("encoder_model_4_block_3_conv_weight"), val = tensor<fp32, [128, 64, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2494912)))];
18
+ tensor<fp32, [256]> encoder_model_6_conv_bias = const()[name = tensor<string, []>("encoder_model_6_conv_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2527744)))];
19
+ tensor<fp32, [256, 128, 10]> encoder_model_6_conv_weight = const()[name = tensor<string, []>("encoder_model_6_conv_weight"), val = tensor<fp32, [256, 128, 10]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2528832)))];
20
+ tensor<fp32, [128]> encoder_model_7_block_1_conv_bias = const()[name = tensor<string, []>("encoder_model_7_block_1_conv_bias"), val = tensor<fp32, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3839616)))];
21
+ tensor<fp32, [128, 256, 3]> encoder_model_7_block_1_conv_weight = const()[name = tensor<string, []>("encoder_model_7_block_1_conv_weight"), val = tensor<fp32, [128, 256, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3840192)))];
22
+ tensor<fp32, [256]> encoder_model_7_block_3_conv_bias = const()[name = tensor<string, []>("encoder_model_7_block_3_conv_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4233472)))];
23
+ tensor<fp32, [256, 128, 1]> encoder_model_7_block_3_conv_weight = const()[name = tensor<string, []>("encoder_model_7_block_3_conv_weight"), val = tensor<fp32, [256, 128, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4234560)))];
24
+ tensor<fp32, [512]> encoder_model_9_conv_bias = const()[name = tensor<string, []>("encoder_model_9_conv_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4365696)))];
25
+ tensor<fp32, [512, 256, 12]> encoder_model_9_conv_weight = const()[name = tensor<string, []>("encoder_model_9_conv_weight"), val = tensor<fp32, [512, 256, 12]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4367808)))];
26
+ tensor<fp32, [512]> encoder_model_11_conv_bias = const()[name = tensor<string, []>("encoder_model_11_conv_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10659328)))];
27
+ tensor<fp32, [512, 512, 3]> encoder_model_11_conv_weight = const()[name = tensor<string, []>("encoder_model_11_conv_weight"), val = tensor<fp32, [512, 512, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10661440)))];
28
+ tensor<fp32, [512]> encoder_transformer_transformer_layers_0_norm1_bias = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_0_norm1_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13807232)))];
29
+ tensor<fp32, [512]> encoder_transformer_transformer_layers_0_norm1_weight = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_0_norm1_weight"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13809344)))];
30
+ tensor<fp32, [1536, 512]> encoder_transformer_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_0_self_attn_in_proj_weight"), val = tensor<fp32, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13811456)))];
31
+ tensor<fp32, [512, 512]> encoder_transformer_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_0_self_attn_out_proj_weight"), val = tensor<fp32, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16957248)))];
32
+ tensor<fp32, [512]> encoder_transformer_transformer_layers_0_layer_scale_1_scale = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_0_layer_scale_1_scale"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18005888)))];
33
+ tensor<fp32, [512]> encoder_transformer_transformer_layers_0_norm2_bias = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_0_norm2_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18008000)))];
34
+ tensor<fp32, [512]> encoder_transformer_transformer_layers_0_norm2_weight = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_0_norm2_weight"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18010112)))];
35
+ tensor<fp32, [2048, 512]> encoder_transformer_transformer_layers_0_linear1_weight = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_0_linear1_weight"), val = tensor<fp32, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18012224)))];
36
+ tensor<fp32, [512, 2048]> encoder_transformer_transformer_layers_0_linear2_weight = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_0_linear2_weight"), val = tensor<fp32, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22206592)))];
37
+ tensor<fp32, [512]> encoder_transformer_transformer_layers_0_layer_scale_2_scale = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_0_layer_scale_2_scale"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26400960)))];
38
+ tensor<fp32, [512]> encoder_transformer_transformer_layers_1_norm1_bias = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_1_norm1_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26403072)))];
39
+ tensor<fp32, [512]> encoder_transformer_transformer_layers_1_norm1_weight = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_1_norm1_weight"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26405184)))];
40
+ tensor<fp32, [1536, 512]> encoder_transformer_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_1_self_attn_in_proj_weight"), val = tensor<fp32, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26407296)))];
41
+ tensor<fp32, [512, 512]> encoder_transformer_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_1_self_attn_out_proj_weight"), val = tensor<fp32, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29553088)))];
42
+ tensor<fp32, [512]> encoder_transformer_transformer_layers_1_layer_scale_1_scale = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_1_layer_scale_1_scale"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30601728)))];
43
+ tensor<fp32, [512]> encoder_transformer_transformer_layers_1_norm2_bias = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_1_norm2_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30603840)))];
44
+ tensor<fp32, [512]> encoder_transformer_transformer_layers_1_norm2_weight = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_1_norm2_weight"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30605952)))];
45
+ tensor<fp32, [2048, 512]> encoder_transformer_transformer_layers_1_linear1_weight = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_1_linear1_weight"), val = tensor<fp32, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30608064)))];
46
+ tensor<fp32, [512, 2048]> encoder_transformer_transformer_layers_1_linear2_weight = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_1_linear2_weight"), val = tensor<fp32, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34802432)))];
47
+ tensor<fp32, [512]> encoder_transformer_transformer_layers_1_layer_scale_2_scale = const()[name = tensor<string, []>("encoder_transformer_transformer_layers_1_layer_scale_2_scale"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38996800)))];
48
+ tensor<fp32, [512, 512, 32]> downsample_conv_conv_weight = const()[name = tensor<string, []>("downsample_conv_conv_weight"), val = tensor<fp32, [512, 512, 32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38998912)))];
49
+ tensor<fp32, []> var_8 = const()[name = tensor<string, []>("op_8"), val = tensor<fp32, []>(0x1p+0)];
50
+ tensor<fp32, []> const_0 = const()[name = tensor<string, []>("const_0"), val = tensor<fp32, []>(0x0p+0)];
51
+ tensor<int32, [6]> input_1_pad_0 = const()[name = tensor<string, []>("input_1_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 6, 0])];
52
+ tensor<string, []> input_1_mode_0 = const()[name = tensor<string, []>("input_1_mode_0"), val = tensor<string, []>("constant")];
53
+ tensor<fp32, [1, 1, ?]> input_1 = pad(constant_val = const_0, mode = input_1_mode_0, pad = input_1_pad_0, x = audio)[name = tensor<string, []>("input_1")];
54
+ tensor<string, []> input_3_pad_type_0 = const()[name = tensor<string, []>("input_3_pad_type_0"), val = tensor<string, []>("valid")];
55
+ tensor<int32, [1]> input_3_strides_0 = const()[name = tensor<string, []>("input_3_strides_0"), val = tensor<int32, [1]>([1])];
56
+ tensor<int32, [2]> input_3_pad_0 = const()[name = tensor<string, []>("input_3_pad_0"), val = tensor<int32, [2]>([0, 0])];
57
+ tensor<int32, [1]> input_3_dilations_0 = const()[name = tensor<string, []>("input_3_dilations_0"), val = tensor<int32, [1]>([1])];
58
+ tensor<int32, []> input_3_groups_0 = const()[name = tensor<string, []>("input_3_groups_0"), val = tensor<int32, []>(1)];
59
+ tensor<fp32, [1, 64, ?]> input_3 = conv(bias = encoder_model_0_conv_bias, dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = encoder_model_0_conv_weight, x = input_1)[name = tensor<string, []>("input_3")];
60
+ tensor<fp32, [1, 64, ?]> input_5 = elu(alpha = var_8, x = input_3)[name = tensor<string, []>("input_5")];
61
+ tensor<fp32, []> const_1 = const()[name = tensor<string, []>("const_1"), val = tensor<fp32, []>(0x0p+0)];
62
+ tensor<int32, [6]> input_7_pad_0 = const()[name = tensor<string, []>("input_7_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 2, 0])];
63
+ tensor<string, []> input_7_mode_0 = const()[name = tensor<string, []>("input_7_mode_0"), val = tensor<string, []>("constant")];
64
+ tensor<fp32, [1, 64, ?]> input_7 = pad(constant_val = const_1, mode = input_7_mode_0, pad = input_7_pad_0, x = input_5)[name = tensor<string, []>("input_7")];
65
+ tensor<string, []> input_9_pad_type_0 = const()[name = tensor<string, []>("input_9_pad_type_0"), val = tensor<string, []>("valid")];
66
+ tensor<int32, [1]> input_9_strides_0 = const()[name = tensor<string, []>("input_9_strides_0"), val = tensor<int32, [1]>([1])];
67
+ tensor<int32, [2]> input_9_pad_0 = const()[name = tensor<string, []>("input_9_pad_0"), val = tensor<int32, [2]>([0, 0])];
68
+ tensor<int32, [1]> input_9_dilations_0 = const()[name = tensor<string, []>("input_9_dilations_0"), val = tensor<int32, [1]>([1])];
69
+ tensor<int32, []> input_9_groups_0 = const()[name = tensor<string, []>("input_9_groups_0"), val = tensor<int32, []>(1)];
70
+ tensor<fp32, [1, 32, ?]> input_9 = conv(bias = encoder_model_1_block_1_conv_bias, dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = encoder_model_1_block_1_conv_weight, x = input_7)[name = tensor<string, []>("input_9")];
71
+ tensor<fp32, [1, 32, ?]> input_11 = elu(alpha = var_8, x = input_9)[name = tensor<string, []>("input_11")];
72
+ tensor<string, []> v_1_pad_type_0 = const()[name = tensor<string, []>("v_1_pad_type_0"), val = tensor<string, []>("valid")];
73
+ tensor<int32, [1]> v_1_strides_0 = const()[name = tensor<string, []>("v_1_strides_0"), val = tensor<int32, [1]>([1])];
74
+ tensor<int32, [2]> v_1_pad_0 = const()[name = tensor<string, []>("v_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
75
+ tensor<int32, [1]> v_1_dilations_0 = const()[name = tensor<string, []>("v_1_dilations_0"), val = tensor<int32, [1]>([1])];
76
+ tensor<int32, []> v_1_groups_0 = const()[name = tensor<string, []>("v_1_groups_0"), val = tensor<int32, []>(1)];
77
+ tensor<fp32, [1, 64, ?]> v_1 = conv(bias = encoder_model_1_block_3_conv_bias, dilations = v_1_dilations_0, groups = v_1_groups_0, pad = v_1_pad_0, pad_type = v_1_pad_type_0, strides = v_1_strides_0, weight = encoder_model_1_block_3_conv_weight, x = input_11)[name = tensor<string, []>("v_1")];
78
+ tensor<fp32, [1, 64, ?]> input_13 = add(x = input_3, y = v_1)[name = tensor<string, []>("input_13")];
79
+ tensor<fp32, [1, 64, ?]> input_15 = elu(alpha = var_8, x = input_13)[name = tensor<string, []>("input_15")];
80
+ tensor<fp32, []> const_2 = const()[name = tensor<string, []>("const_2"), val = tensor<fp32, []>(0x0p+0)];
81
+ tensor<int32, [6]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 4, 0])];
82
+ tensor<string, []> input_17_mode_0 = const()[name = tensor<string, []>("input_17_mode_0"), val = tensor<string, []>("constant")];
83
+ tensor<fp32, [1, 64, ?]> input_17 = pad(constant_val = const_2, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15)[name = tensor<string, []>("input_17")];
84
+ tensor<string, []> input_19_pad_type_0 = const()[name = tensor<string, []>("input_19_pad_type_0"), val = tensor<string, []>("valid")];
85
+ tensor<int32, [1]> input_19_strides_0 = const()[name = tensor<string, []>("input_19_strides_0"), val = tensor<int32, [1]>([4])];
86
+ tensor<int32, [2]> input_19_pad_0 = const()[name = tensor<string, []>("input_19_pad_0"), val = tensor<int32, [2]>([0, 0])];
87
+ tensor<int32, [1]> input_19_dilations_0 = const()[name = tensor<string, []>("input_19_dilations_0"), val = tensor<int32, [1]>([1])];
88
+ tensor<int32, []> input_19_groups_0 = const()[name = tensor<string, []>("input_19_groups_0"), val = tensor<int32, []>(1)];
89
+ tensor<fp32, [1, 128, ?]> input_19 = conv(bias = encoder_model_3_conv_bias, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = encoder_model_3_conv_weight, x = input_17)[name = tensor<string, []>("input_19")];
90
+ tensor<fp32, [1, 128, ?]> input_21 = elu(alpha = var_8, x = input_19)[name = tensor<string, []>("input_21")];
91
+ tensor<fp32, []> const_3 = const()[name = tensor<string, []>("const_3"), val = tensor<fp32, []>(0x0p+0)];
92
+ tensor<int32, [6]> input_23_pad_0 = const()[name = tensor<string, []>("input_23_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 2, 0])];
93
+ tensor<string, []> input_23_mode_0 = const()[name = tensor<string, []>("input_23_mode_0"), val = tensor<string, []>("constant")];
94
+ tensor<fp32, [1, 128, ?]> input_23 = pad(constant_val = const_3, mode = input_23_mode_0, pad = input_23_pad_0, x = input_21)[name = tensor<string, []>("input_23")];
95
+ tensor<string, []> input_25_pad_type_0 = const()[name = tensor<string, []>("input_25_pad_type_0"), val = tensor<string, []>("valid")];
96
+ tensor<int32, [1]> input_25_strides_0 = const()[name = tensor<string, []>("input_25_strides_0"), val = tensor<int32, [1]>([1])];
97
+ tensor<int32, [2]> input_25_pad_0 = const()[name = tensor<string, []>("input_25_pad_0"), val = tensor<int32, [2]>([0, 0])];
98
+ tensor<int32, [1]> input_25_dilations_0 = const()[name = tensor<string, []>("input_25_dilations_0"), val = tensor<int32, [1]>([1])];
99
+ tensor<int32, []> input_25_groups_0 = const()[name = tensor<string, []>("input_25_groups_0"), val = tensor<int32, []>(1)];
100
+ tensor<fp32, [1, 64, ?]> input_25 = conv(bias = encoder_model_4_block_1_conv_bias, dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = encoder_model_4_block_1_conv_weight, x = input_23)[name = tensor<string, []>("input_25")];
101
+ tensor<fp32, [1, 64, ?]> input_27 = elu(alpha = var_8, x = input_25)[name = tensor<string, []>("input_27")];
102
+ tensor<string, []> v_3_pad_type_0 = const()[name = tensor<string, []>("v_3_pad_type_0"), val = tensor<string, []>("valid")];
103
+ tensor<int32, [1]> v_3_strides_0 = const()[name = tensor<string, []>("v_3_strides_0"), val = tensor<int32, [1]>([1])];
104
+ tensor<int32, [2]> v_3_pad_0 = const()[name = tensor<string, []>("v_3_pad_0"), val = tensor<int32, [2]>([0, 0])];
105
+ tensor<int32, [1]> v_3_dilations_0 = const()[name = tensor<string, []>("v_3_dilations_0"), val = tensor<int32, [1]>([1])];
106
+ tensor<int32, []> v_3_groups_0 = const()[name = tensor<string, []>("v_3_groups_0"), val = tensor<int32, []>(1)];
107
+ tensor<fp32, [1, 128, ?]> v_3 = conv(bias = encoder_model_4_block_3_conv_bias, dilations = v_3_dilations_0, groups = v_3_groups_0, pad = v_3_pad_0, pad_type = v_3_pad_type_0, strides = v_3_strides_0, weight = encoder_model_4_block_3_conv_weight, x = input_27)[name = tensor<string, []>("v_3")];
108
+ tensor<fp32, [1, 128, ?]> input_29 = add(x = input_19, y = v_3)[name = tensor<string, []>("input_29")];
109
+ tensor<fp32, [1, 128, ?]> input_31 = elu(alpha = var_8, x = input_29)[name = tensor<string, []>("input_31")];
110
+ tensor<fp32, []> const_4 = const()[name = tensor<string, []>("const_4"), val = tensor<fp32, []>(0x0p+0)];
111
+ tensor<int32, [6]> input_33_pad_0 = const()[name = tensor<string, []>("input_33_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 5, 0])];
112
+ tensor<string, []> input_33_mode_0 = const()[name = tensor<string, []>("input_33_mode_0"), val = tensor<string, []>("constant")];
113
+ tensor<fp32, [1, 128, ?]> input_33 = pad(constant_val = const_4, mode = input_33_mode_0, pad = input_33_pad_0, x = input_31)[name = tensor<string, []>("input_33")];
114
+ tensor<string, []> input_35_pad_type_0 = const()[name = tensor<string, []>("input_35_pad_type_0"), val = tensor<string, []>("valid")];
115
+ tensor<int32, [1]> input_35_strides_0 = const()[name = tensor<string, []>("input_35_strides_0"), val = tensor<int32, [1]>([5])];
116
+ tensor<int32, [2]> input_35_pad_0 = const()[name = tensor<string, []>("input_35_pad_0"), val = tensor<int32, [2]>([0, 0])];
117
+ tensor<int32, [1]> input_35_dilations_0 = const()[name = tensor<string, []>("input_35_dilations_0"), val = tensor<int32, [1]>([1])];
118
+ tensor<int32, []> input_35_groups_0 = const()[name = tensor<string, []>("input_35_groups_0"), val = tensor<int32, []>(1)];
119
+ tensor<fp32, [1, 256, ?]> input_35 = conv(bias = encoder_model_6_conv_bias, dilations = input_35_dilations_0, groups = input_35_groups_0, pad = input_35_pad_0, pad_type = input_35_pad_type_0, strides = input_35_strides_0, weight = encoder_model_6_conv_weight, x = input_33)[name = tensor<string, []>("input_35")];
120
+ tensor<fp32, [1, 256, ?]> input_37 = elu(alpha = var_8, x = input_35)[name = tensor<string, []>("input_37")];
121
+ tensor<fp32, []> const_5 = const()[name = tensor<string, []>("const_5"), val = tensor<fp32, []>(0x0p+0)];
122
+ tensor<int32, [6]> input_39_pad_0 = const()[name = tensor<string, []>("input_39_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 2, 0])];
123
+ tensor<string, []> input_39_mode_0 = const()[name = tensor<string, []>("input_39_mode_0"), val = tensor<string, []>("constant")];
124
+ tensor<fp32, [1, 256, ?]> input_39 = pad(constant_val = const_5, mode = input_39_mode_0, pad = input_39_pad_0, x = input_37)[name = tensor<string, []>("input_39")];
125
+ tensor<string, []> input_41_pad_type_0 = const()[name = tensor<string, []>("input_41_pad_type_0"), val = tensor<string, []>("valid")];
126
+ tensor<int32, [1]> input_41_strides_0 = const()[name = tensor<string, []>("input_41_strides_0"), val = tensor<int32, [1]>([1])];
127
+ tensor<int32, [2]> input_41_pad_0 = const()[name = tensor<string, []>("input_41_pad_0"), val = tensor<int32, [2]>([0, 0])];
128
+ tensor<int32, [1]> input_41_dilations_0 = const()[name = tensor<string, []>("input_41_dilations_0"), val = tensor<int32, [1]>([1])];
129
+ tensor<int32, []> input_41_groups_0 = const()[name = tensor<string, []>("input_41_groups_0"), val = tensor<int32, []>(1)];
130
+ tensor<fp32, [1, 128, ?]> input_41 = conv(bias = encoder_model_7_block_1_conv_bias, dilations = input_41_dilations_0, groups = input_41_groups_0, pad = input_41_pad_0, pad_type = input_41_pad_type_0, strides = input_41_strides_0, weight = encoder_model_7_block_1_conv_weight, x = input_39)[name = tensor<string, []>("input_41")];
131
+ tensor<fp32, [1, 128, ?]> input_43 = elu(alpha = var_8, x = input_41)[name = tensor<string, []>("input_43")];
132
+ tensor<string, []> v_pad_type_0 = const()[name = tensor<string, []>("v_pad_type_0"), val = tensor<string, []>("valid")];
133
+ tensor<int32, [1]> v_strides_0 = const()[name = tensor<string, []>("v_strides_0"), val = tensor<int32, [1]>([1])];
134
+ tensor<int32, [2]> v_pad_0 = const()[name = tensor<string, []>("v_pad_0"), val = tensor<int32, [2]>([0, 0])];
135
+ tensor<int32, [1]> v_dilations_0 = const()[name = tensor<string, []>("v_dilations_0"), val = tensor<int32, [1]>([1])];
136
+ tensor<int32, []> v_groups_0 = const()[name = tensor<string, []>("v_groups_0"), val = tensor<int32, []>(1)];
137
+ tensor<fp32, [1, 256, ?]> v = conv(bias = encoder_model_7_block_3_conv_bias, dilations = v_dilations_0, groups = v_groups_0, pad = v_pad_0, pad_type = v_pad_type_0, strides = v_strides_0, weight = encoder_model_7_block_3_conv_weight, x = input_43)[name = tensor<string, []>("v")];
138
+ tensor<fp32, [1, 256, ?]> input_45 = add(x = input_35, y = v)[name = tensor<string, []>("input_45")];
139
+ tensor<fp32, [1, 256, ?]> input_47 = elu(alpha = var_8, x = input_45)[name = tensor<string, []>("input_47")];
140
+ tensor<fp32, []> const_6 = const()[name = tensor<string, []>("const_6"), val = tensor<fp32, []>(0x0p+0)];
141
+ tensor<int32, [6]> input_49_pad_0 = const()[name = tensor<string, []>("input_49_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 6, 0])];
142
+ tensor<string, []> input_49_mode_0 = const()[name = tensor<string, []>("input_49_mode_0"), val = tensor<string, []>("constant")];
143
+ tensor<fp32, [1, 256, ?]> input_49 = pad(constant_val = const_6, mode = input_49_mode_0, pad = input_49_pad_0, x = input_47)[name = tensor<string, []>("input_49")];
144
+ tensor<string, []> input_51_pad_type_0 = const()[name = tensor<string, []>("input_51_pad_type_0"), val = tensor<string, []>("valid")];
145
+ tensor<int32, [1]> input_51_strides_0 = const()[name = tensor<string, []>("input_51_strides_0"), val = tensor<int32, [1]>([6])];
146
+ tensor<int32, [2]> input_51_pad_0 = const()[name = tensor<string, []>("input_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
147
+ tensor<int32, [1]> input_51_dilations_0 = const()[name = tensor<string, []>("input_51_dilations_0"), val = tensor<int32, [1]>([1])];
148
+ tensor<int32, []> input_51_groups_0 = const()[name = tensor<string, []>("input_51_groups_0"), val = tensor<int32, []>(1)];
149
+ tensor<fp32, [1, 512, ?]> input_51 = conv(bias = encoder_model_9_conv_bias, dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = encoder_model_9_conv_weight, x = input_49)[name = tensor<string, []>("input_51")];
150
+ tensor<fp32, [1, 512, ?]> input_53 = elu(alpha = var_8, x = input_51)[name = tensor<string, []>("input_53")];
151
+ tensor<fp32, []> const_7 = const()[name = tensor<string, []>("const_7"), val = tensor<fp32, []>(0x0p+0)];
152
+ tensor<int32, [6]> input_55_pad_0 = const()[name = tensor<string, []>("input_55_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 2, 0])];
153
+ tensor<string, []> input_55_mode_0 = const()[name = tensor<string, []>("input_55_mode_0"), val = tensor<string, []>("constant")];
154
+ tensor<fp32, [1, 512, ?]> input_55 = pad(constant_val = const_7, mode = input_55_mode_0, pad = input_55_pad_0, x = input_53)[name = tensor<string, []>("input_55")];
155
+ tensor<string, []> x_1_pad_type_0 = const()[name = tensor<string, []>("x_1_pad_type_0"), val = tensor<string, []>("valid")];
156
+ tensor<int32, [1]> x_1_strides_0 = const()[name = tensor<string, []>("x_1_strides_0"), val = tensor<int32, [1]>([1])];
157
+ tensor<int32, [2]> x_1_pad_0 = const()[name = tensor<string, []>("x_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
158
+ tensor<int32, [1]> x_1_dilations_0 = const()[name = tensor<string, []>("x_1_dilations_0"), val = tensor<int32, [1]>([1])];
159
+ tensor<int32, []> x_1_groups_0 = const()[name = tensor<string, []>("x_1_groups_0"), val = tensor<int32, []>(1)];
160
+ tensor<fp32, [1, 512, ?]> x_1 = conv(bias = encoder_model_11_conv_bias, dilations = x_1_dilations_0, groups = x_1_groups_0, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = x_1_strides_0, weight = encoder_model_11_conv_weight, x = input_55)[name = tensor<string, []>("x_1")];
161
+ tensor<int32, []> var_163 = const()[name = tensor<string, []>("op_163"), val = tensor<int32, []>(0)];
162
+ tensor<int32, []> var_169 = const()[name = tensor<string, []>("op_169"), val = tensor<int32, []>(3)];
163
+ tensor<int32, []> var_170 = const()[name = tensor<string, []>("op_170"), val = tensor<int32, []>(8)];
164
+ tensor<int32, []> var_171 = const()[name = tensor<string, []>("op_171"), val = tensor<int32, []>(-1)];
165
+ tensor<int32, []> var_175 = const()[name = tensor<string, []>("op_175"), val = tensor<int32, []>(250)];
166
+ tensor<fp32, []> var_178 = const()[name = tensor<string, []>("op_178"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
167
+ tensor<int32, []> var_180 = const()[name = tensor<string, []>("op_180"), val = tensor<int32, []>(2)];
168
+ tensor<int32, [3]> input_57_perm_0 = const()[name = tensor<string, []>("input_57_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
169
+ tensor<int32, [1]> query_1_axes_0 = const()[name = tensor<string, []>("query_1_axes_0"), val = tensor<int32, [1]>([-1])];
170
+ tensor<fp32, [1, ?, 512]> input_57 = transpose(perm = input_57_perm_0, x = x_1)[name = tensor<string, []>("transpose_14")];
171
+ tensor<fp32, [1, ?, 512]> query_1 = layer_norm(axes = query_1_axes_0, beta = encoder_transformer_transformer_layers_0_norm1_bias, epsilon = var_178, gamma = encoder_transformer_transformer_layers_0_norm1_weight, x = input_57)[name = tensor<string, []>("query_1")];
172
+ tensor<int32, [3]> var_202_shape = shape(x = query_1)[name = tensor<string, []>("op_202_shape")];
173
+ tensor<int32, []> gather_1_batch_dims_0 = const()[name = tensor<string, []>("gather_1_batch_dims_0"), val = tensor<int32, []>(0)];
174
+ tensor<bool, []> gather_1_validate_indices_0 = const()[name = tensor<string, []>("gather_1_validate_indices_0"), val = tensor<bool, []>(false)];
175
+ tensor<int32, []> select_3 = const()[name = tensor<string, []>("select_3"), val = tensor<int32, []>(1)];
176
+ tensor<int32, []> gather_1_axis_1 = const()[name = tensor<string, []>("gather_1_axis_1"), val = tensor<int32, []>(0)];
177
+ tensor<int32, []> gather_1 = gather(axis = gather_1_axis_1, batch_dims = gather_1_batch_dims_0, indices = select_3, validate_indices = gather_1_validate_indices_0, x = var_202_shape)[name = tensor<string, []>("gather_1")];
178
+ tensor<fp32, [1536]> linear_0_bias_0 = const()[name = tensor<string, []>("linear_0_bias_0"), val = tensor<fp32, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72553408)))];
179
+ tensor<fp32, [1, ?, 1536]> x_3 = linear(bias = linear_0_bias_0, weight = encoder_transformer_transformer_layers_0_self_attn_in_proj_weight, x = query_1)[name = tensor<string, []>("linear_0")];
180
+ tensor<int32, [5]> concat_0x = const()[name = tensor<string, []>("concat_0x"), val = tensor<int32, [5]>([1, -1, 3, 8, 64])];
181
+ tensor<fp32, [1, ?, 3, 8, 64]> x_5 = reshape(shape = concat_0x, x = x_3)[name = tensor<string, []>("x_5")];
182
+ tensor<int32, [5]> var_215 = const()[name = tensor<string, []>("op_215"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
183
+ tensor<int32, [3]> var_217_split_sizes_0 = const()[name = tensor<string, []>("op_217_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
184
+ tensor<int32, []> var_217_axis_0 = const()[name = tensor<string, []>("op_217_axis_0"), val = tensor<int32, []>(0)];
185
+ tensor<fp32, [3, 1, 8, ?, 64]> var_216 = transpose(perm = var_215, x = x_5)[name = tensor<string, []>("transpose_13")];
186
+ tensor<fp32, [1, 1, 8, ?, 64]> var_217_0, tensor<fp32, [1, 1, 8, ?, 64]> var_217_1, tensor<fp32, [1, 1, 8, ?, 64]> var_217_2 = split(axis = var_217_axis_0, split_sizes = var_217_split_sizes_0, x = var_216)[name = tensor<string, []>("op_217")];
187
+ tensor<int32, [1]> squeeze_0_axes_0 = const()[name = tensor<string, []>("squeeze_0_axes_0"), val = tensor<int32, [1]>([0])];
188
+ tensor<fp32, [1, 8, ?, 64]> squeeze_0 = squeeze(axes = squeeze_0_axes_0, x = var_217_0)[name = tensor<string, []>("squeeze_0")];
189
+ tensor<int32, [1]> squeeze_1_axes_0 = const()[name = tensor<string, []>("squeeze_1_axes_0"), val = tensor<int32, [1]>([0])];
190
+ tensor<fp32, [1, 8, ?, 64]> squeeze_1 = squeeze(axes = squeeze_1_axes_0, x = var_217_1)[name = tensor<string, []>("squeeze_1")];
191
+ tensor<int32, [1]> squeeze_2_axes_0 = const()[name = tensor<string, []>("squeeze_2_axes_0"), val = tensor<int32, [1]>([0])];
192
+ tensor<fp32, [1, 8, ?, 64]> squeeze_2 = squeeze(axes = squeeze_2_axes_0, x = var_217_2)[name = tensor<string, []>("squeeze_2")];
193
+ tensor<int32, [4]> var_221 = const()[name = tensor<string, []>("op_221"), val = tensor<int32, [4]>([0, 2, 1, 3])];
194
+ tensor<int32, [4]> var_223 = const()[name = tensor<string, []>("op_223"), val = tensor<int32, [4]>([0, 2, 1, 3])];
195
+ tensor<fp32, [1, ?, 8, 64]> q_3 = transpose(perm = var_221, x = squeeze_0)[name = tensor<string, []>("transpose_12")];
196
+ tensor<int32, [4]> var_225_shape = shape(x = q_3)[name = tensor<string, []>("op_225_shape")];
197
+ tensor<int32, []> gather_6_batch_dims_0 = const()[name = tensor<string, []>("gather_6_batch_dims_0"), val = tensor<int32, []>(0)];
198
+ tensor<bool, []> gather_6_validate_indices_0 = const()[name = tensor<string, []>("gather_6_validate_indices_0"), val = tensor<bool, []>(false)];
199
+ tensor<int32, []> select_5 = const()[name = tensor<string, []>("select_5"), val = tensor<int32, []>(1)];
200
+ tensor<int32, []> gather_6_axis_1 = const()[name = tensor<string, []>("gather_6_axis_1"), val = tensor<int32, []>(0)];
201
+ 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_225_shape)[name = tensor<string, []>("gather_6")];
202
+ tensor<fp32, [32]> freqs_1 = const()[name = tensor<string, []>("freqs_1"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72559616)))];
203
+ tensor<int32, []> const_10 = const()[name = tensor<string, []>("const_10"), val = tensor<int32, []>(0)];
204
+ tensor<int32, []> const_11 = const()[name = tensor<string, []>("const_11"), val = tensor<int32, []>(1)];
205
+ tensor<int32, [?]> ts_1 = range_1d(end = gather_6, start = const_10, step = const_11)[name = tensor<string, []>("ts_1")];
206
+ tensor<int32, [3]> var_239 = const()[name = tensor<string, []>("op_239"), val = tensor<int32, [3]>([-1, 1, 1])];
207
+ tensor<int32, [?, 1, 1]> ts_5 = reshape(shape = var_239, x = ts_1)[name = tensor<string, []>("ts_5")];
208
+ tensor<int32, [5]> concat_1x = const()[name = tensor<string, []>("concat_1x"), val = tensor<int32, [5]>([1, -1, 8, 32, 2])];
209
+ tensor<fp32, [1, ?, 8, 32, 2]> q_5 = reshape(shape = concat_1x, x = q_3)[name = tensor<string, []>("q_5")];
210
+ tensor<int32, [5]> concat_2x = const()[name = tensor<string, []>("concat_2x"), val = tensor<int32, [5]>([1, -1, 8, 32, 2])];
211
+ tensor<fp32, [1, ?, 8, 64]> k_3 = transpose(perm = var_223, x = squeeze_1)[name = tensor<string, []>("transpose_11")];
212
+ tensor<fp32, [1, ?, 8, 32, 2]> k_5 = reshape(shape = concat_2x, x = k_3)[name = tensor<string, []>("k_5")];
213
+ tensor<int32, [5]> var_249_begin_0 = const()[name = tensor<string, []>("op_249_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
214
+ tensor<int32, [5]> var_249_end_0 = const()[name = tensor<string, []>("op_249_end_0"), val = tensor<int32, [5]>([1, 0, 8, 32, 1])];
215
+ tensor<bool, [5]> var_249_end_mask_0 = const()[name = tensor<string, []>("op_249_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
216
+ tensor<bool, [5]> var_249_squeeze_mask_0 = const()[name = tensor<string, []>("op_249_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
217
+ tensor<fp32, [1, ?, 8, 32]> var_249 = slice_by_index(begin = var_249_begin_0, end = var_249_end_0, end_mask = var_249_end_mask_0, squeeze_mask = var_249_squeeze_mask_0, x = q_5)[name = tensor<string, []>("op_249")];
218
+ tensor<int32, [5]> var_251_begin_0 = const()[name = tensor<string, []>("op_251_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
219
+ tensor<int32, [5]> var_251_end_0 = const()[name = tensor<string, []>("op_251_end_0"), val = tensor<int32, [5]>([1, 0, 8, 32, 2])];
220
+ tensor<bool, [5]> var_251_end_mask_0 = const()[name = tensor<string, []>("op_251_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
221
+ tensor<bool, [5]> var_251_squeeze_mask_0 = const()[name = tensor<string, []>("op_251_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
222
+ tensor<fp32, [1, ?, 8, 32]> var_251 = slice_by_index(begin = var_251_begin_0, end = var_251_end_0, end_mask = var_251_end_mask_0, squeeze_mask = var_251_squeeze_mask_0, x = q_5)[name = tensor<string, []>("op_251")];
223
+ tensor<int32, [5]> var_253_begin_0 = const()[name = tensor<string, []>("op_253_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
224
+ tensor<int32, [5]> var_253_end_0 = const()[name = tensor<string, []>("op_253_end_0"), val = tensor<int32, [5]>([1, 0, 8, 32, 1])];
225
+ tensor<bool, [5]> var_253_end_mask_0 = const()[name = tensor<string, []>("op_253_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
226
+ tensor<bool, [5]> var_253_squeeze_mask_0 = const()[name = tensor<string, []>("op_253_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
227
+ tensor<fp32, [1, ?, 8, 32]> var_253 = slice_by_index(begin = var_253_begin_0, end = var_253_end_0, end_mask = var_253_end_mask_0, squeeze_mask = var_253_squeeze_mask_0, x = k_5)[name = tensor<string, []>("op_253")];
228
+ tensor<int32, [5]> var_255_begin_0 = const()[name = tensor<string, []>("op_255_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
229
+ tensor<int32, [5]> var_255_end_0 = const()[name = tensor<string, []>("op_255_end_0"), val = tensor<int32, [5]>([1, 0, 8, 32, 2])];
230
+ tensor<bool, [5]> var_255_end_mask_0 = const()[name = tensor<string, []>("op_255_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
231
+ tensor<bool, [5]> var_255_squeeze_mask_0 = const()[name = tensor<string, []>("op_255_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
232
+ tensor<fp32, [1, ?, 8, 32]> var_255 = slice_by_index(begin = var_255_begin_0, end = var_255_end_0, end_mask = var_255_end_mask_0, squeeze_mask = var_255_squeeze_mask_0, x = k_5)[name = tensor<string, []>("op_255")];
233
+ tensor<string, []> ts_5_promoted_dtype_0 = const()[name = tensor<string, []>("ts_5_promoted_dtype_0"), val = tensor<string, []>("fp32")];
234
+ tensor<fp32, [?, 1, 1]> ts_5_promoted = cast(dtype = ts_5_promoted_dtype_0, x = ts_5)[name = tensor<string, []>("cast_34")];
235
+ tensor<fp32, [?, 1, 32]> var_257 = mul(x = freqs_1, y = ts_5_promoted)[name = tensor<string, []>("op_257")];
236
+ tensor<fp32, [?, 1, 32]> rotr_1 = cos(x = var_257)[name = tensor<string, []>("rotr_1")];
237
+ tensor<fp32, [?, 1, 32]> roti_1 = sin(x = var_257)[name = tensor<string, []>("roti_1")];
238
+ tensor<fp32, [1, ?, 8, 32]> var_261 = mul(x = var_249, y = rotr_1)[name = tensor<string, []>("op_261")];
239
+ tensor<fp32, [1, ?, 8, 32]> var_262 = mul(x = var_251, y = roti_1)[name = tensor<string, []>("op_262")];
240
+ tensor<fp32, [1, ?, 8, 32]> qor_1 = sub(x = var_261, y = var_262)[name = tensor<string, []>("qor_1")];
241
+ tensor<fp32, [1, ?, 8, 32]> var_264 = mul(x = var_249, y = roti_1)[name = tensor<string, []>("op_264")];
242
+ tensor<fp32, [1, ?, 8, 32]> var_265 = mul(x = var_251, y = rotr_1)[name = tensor<string, []>("op_265")];
243
+ tensor<fp32, [1, ?, 8, 32]> qoi_1 = add(x = var_264, y = var_265)[name = tensor<string, []>("qoi_1")];
244
+ tensor<fp32, [1, ?, 8, 32]> var_267 = mul(x = var_253, y = rotr_1)[name = tensor<string, []>("op_267")];
245
+ tensor<fp32, [1, ?, 8, 32]> var_268 = mul(x = var_255, y = roti_1)[name = tensor<string, []>("op_268")];
246
+ tensor<fp32, [1, ?, 8, 32]> kor_1 = sub(x = var_267, y = var_268)[name = tensor<string, []>("kor_1")];
247
+ tensor<fp32, [1, ?, 8, 32]> var_270 = mul(x = var_253, y = roti_1)[name = tensor<string, []>("op_270")];
248
+ tensor<fp32, [1, ?, 8, 32]> var_271 = mul(x = var_255, y = rotr_1)[name = tensor<string, []>("op_271")];
249
+ tensor<fp32, [1, ?, 8, 32]> koi_1 = add(x = var_270, y = var_271)[name = tensor<string, []>("koi_1")];
250
+ tensor<int32, []> qo_1_axis_0 = const()[name = tensor<string, []>("qo_1_axis_0"), val = tensor<int32, []>(-1)];
251
+ tensor<fp32, [1, ?, 8, 32, 2]> qo_1 = stack(axis = qo_1_axis_0, values = (qor_1, qoi_1))[name = tensor<string, []>("qo_1")];
252
+ tensor<int32, []> ko_1_axis_0 = const()[name = tensor<string, []>("ko_1_axis_0"), val = tensor<int32, []>(-1)];
253
+ tensor<fp32, [1, ?, 8, 32, 2]> ko_1 = stack(axis = ko_1_axis_0, values = (kor_1, koi_1))[name = tensor<string, []>("ko_1")];
254
+ tensor<int32, [4]> concat_3x = const()[name = tensor<string, []>("concat_3x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
255
+ tensor<fp32, [1, ?, 8, 64]> q_7 = reshape(shape = concat_3x, x = qo_1)[name = tensor<string, []>("q_7")];
256
+ tensor<int32, [4]> concat_4x = const()[name = tensor<string, []>("concat_4x"), val = tensor<int32, [4]>([1, -1, 8, 64])];
257
+ tensor<fp32, [1, ?, 8, 64]> k_7 = reshape(shape = concat_4x, x = ko_1)[name = tensor<string, []>("k_7")];
258
+ tensor<int32, [4]> var_288 = const()[name = tensor<string, []>("op_288"), val = tensor<int32, [4]>([0, 2, 1, 3])];
259
+ tensor<int32, [4]> var_290 = const()[name = tensor<string, []>("op_290"), val = tensor<int32, [4]>([0, 2, 1, 3])];
260
+ tensor<fp32, [1, 8, ?, 64]> keys_1 = transpose(perm = var_290, x = k_7)[name = tensor<string, []>("transpose_9")];
261
+ tensor<int32, [4]> var_293_shape = shape(x = keys_1)[name = tensor<string, []>("op_293_shape")];
262
+ tensor<int32, []> gather_11_batch_dims_0 = const()[name = tensor<string, []>("gather_11_batch_dims_0"), val = tensor<int32, []>(0)];
263
+ tensor<bool, []> gather_11_validate_indices_0 = const()[name = tensor<string, []>("gather_11_validate_indices_0"), val = tensor<bool, []>(false)];
264
+ tensor<int32, []> select_6 = const()[name = tensor<string, []>("select_6"), val = tensor<int32, []>(2)];
265
+ tensor<int32, []> gather_11_axis_1 = const()[name = tensor<string, []>("gather_11_axis_1"), val = tensor<int32, []>(0)];
266
+ tensor<int32, []> gather_11 = gather(axis = gather_11_axis_1, batch_dims = gather_11_batch_dims_0, indices = select_6, validate_indices = gather_11_validate_indices_0, x = var_293_shape)[name = tensor<string, []>("gather_11")];
267
+ tensor<int32, []> const_12 = const()[name = tensor<string, []>("const_12"), val = tensor<int32, []>(0)];
268
+ tensor<int32, []> const_13 = const()[name = tensor<string, []>("const_13"), val = tensor<int32, []>(1)];
269
+ tensor<int32, [?]> positions_1 = range_1d(end = gather_11, start = const_12, step = const_13)[name = tensor<string, []>("positions_1")];
270
+ tensor<int32, [2]> concat_5 = const()[name = tensor<string, []>("concat_5"), val = tensor<int32, [2]>([1, -1])];
271
+ tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = tensor<string, []>("expand_dims_0_axes_0"), val = tensor<int32, [1]>([0])];
272
+ tensor<int32, [1, ?]> expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = positions_1)[name = tensor<string, []>("expand_dims_0")];
273
+ tensor<int32, [2]> shape_0 = shape(x = expand_dims_0)[name = tensor<string, []>("shape_0")];
274
+ tensor<bool, [2]> equal_0 = const()[name = tensor<string, []>("equal_0"), val = tensor<bool, [2]>([false, true])];
275
+ tensor<int32, [2]> select_0 = select(a = shape_0, b = concat_5, cond = equal_0)[name = tensor<string, []>("select_0")];
276
+ tensor<int32, [2]> real_div_0 = real_div(x = select_0, y = shape_0)[name = tensor<string, []>("real_div_0")];
277
+ tensor<int32, [?, ?]> pos_k_1 = tile(reps = real_div_0, x = expand_dims_0)[name = tensor<string, []>("pos_k_1")];
278
+ tensor<int32, [1]> pos_k_3_axes_0 = const()[name = tensor<string, []>("pos_k_3_axes_0"), val = tensor<int32, [1]>([1])];
279
+ tensor<int32, [?, 1, ?]> pos_k_3 = expand_dims(axes = pos_k_3_axes_0, x = pos_k_1)[name = tensor<string, []>("pos_k_3")];
280
+ tensor<int32, [1, 1, 1]> var_300 = const()[name = tensor<string, []>("op_300"), val = tensor<int32, [1, 1, 1]>([[[0]]])];
281
+ tensor<int32, []> const_14 = const()[name = tensor<string, []>("const_14"), val = tensor<int32, []>(0)];
282
+ tensor<int32, []> const_15 = const()[name = tensor<string, []>("const_15"), val = tensor<int32, []>(1)];
283
+ tensor<int32, [?]> var_301 = range_1d(end = gather_1, start = const_14, step = const_15)[name = tensor<string, []>("op_301")];
284
+ tensor<int32, [2]> var_302 = const()[name = tensor<string, []>("op_302"), val = tensor<int32, [2]>([-1, 1])];
285
+ tensor<int32, [?, 1]> var_303 = reshape(shape = var_302, x = var_301)[name = tensor<string, []>("op_303")];
286
+ tensor<int32, [1, ?, 1]> pos_q_1 = add(x = var_300, y = var_303)[name = tensor<string, []>("pos_q_1")];
287
+ tensor<int32, [?, ?, ?]> delta_1 = sub(x = pos_q_1, y = pos_k_3)[name = tensor<string, []>("delta_1")];
288
+ tensor<bool, [?, 1, ?]> var_306 = greater_equal(x = pos_k_3, y = var_163)[name = tensor<string, []>("op_306")];
289
+ tensor<bool, [?, ?, ?]> var_307 = greater_equal(x = delta_1, y = var_163)[name = tensor<string, []>("op_307")];
290
+ tensor<bool, [?, ?, ?]> attn_bias_1 = logical_and(x = var_306, y = var_307)[name = tensor<string, []>("attn_bias_1")];
291
+ tensor<bool, [?, ?, ?]> var_309 = less(x = delta_1, y = var_175)[name = tensor<string, []>("op_309")];
292
+ tensor<bool, [?, ?, ?]> attn_bias_3 = logical_and(x = attn_bias_1, y = var_309)[name = tensor<string, []>("attn_bias_3")];
293
+ tensor<int32, [1]> attn_bias_5_axes_0 = const()[name = tensor<string, []>("attn_bias_5_axes_0"), val = tensor<int32, [1]>([1])];
294
+ tensor<bool, [?, 1, ?, ?]> attn_bias_5 = expand_dims(axes = attn_bias_5_axes_0, x = attn_bias_3)[name = tensor<string, []>("attn_bias_5")];
295
+ tensor<string, []> cast_13_dtype_0 = const()[name = tensor<string, []>("cast_13_dtype_0"), val = tensor<string, []>("fp32")];
296
+ tensor<fp32, []> sub_0_x_0 = const()[name = tensor<string, []>("sub_0_x_0"), val = tensor<fp32, []>(0x1p+0)];
297
+ tensor<fp32, [?, 1, ?, ?]> cast_13 = cast(dtype = cast_13_dtype_0, x = attn_bias_5)[name = tensor<string, []>("cast_33")];
298
+ tensor<fp32, [?, 1, ?, ?]> sub_0 = sub(x = sub_0_x_0, y = cast_13)[name = tensor<string, []>("sub_0")];
299
+ tensor<fp32, []> mul_0_x_0 = const()[name = tensor<string, []>("mul_0_x_0"), val = tensor<fp32, []>(-0x1.d4cp+14)];
300
+ tensor<fp32, [?, 1, ?, ?]> mul_0 = mul(x = mul_0_x_0, y = sub_0)[name = tensor<string, []>("mul_0")];
301
+ tensor<fp32, []> mul_1_y_0 = const()[name = tensor<string, []>("mul_1_y_0"), val = tensor<fp32, []>(0x1p-3)];
302
+ tensor<fp32, [1, 8, ?, 64]> q_9 = transpose(perm = var_288, x = q_7)[name = tensor<string, []>("transpose_10")];
303
+ tensor<fp32, [1, 8, ?, 64]> mul_1 = mul(x = q_9, y = mul_1_y_0)[name = tensor<string, []>("mul_1")];
304
+ tensor<bool, []> matmul_0_transpose_y_0 = const()[name = tensor<string, []>("matmul_0_transpose_y_0"), val = tensor<bool, []>(true)];
305
+ tensor<bool, []> matmul_0_transpose_x_0 = const()[name = tensor<string, []>("matmul_0_transpose_x_0"), val = tensor<bool, []>(false)];
306
+ tensor<fp32, [1, 8, ?, ?]> matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = keys_1)[name = tensor<string, []>("matmul_0")];
307
+ tensor<fp32, [?, 8, ?, ?]> add_0 = add(x = matmul_0, y = mul_0)[name = tensor<string, []>("add_0")];
308
+ tensor<int32, []> softmax_0_axis_0 = const()[name = tensor<string, []>("softmax_0_axis_0"), val = tensor<int32, []>(-1)];
309
+ tensor<fp32, [?, 8, ?, ?]> softmax_0 = softmax(axis = softmax_0_axis_0, x = add_0)[name = tensor<string, []>("softmax_0")];
310
+ tensor<bool, []> x_7_transpose_x_0 = const()[name = tensor<string, []>("x_7_transpose_x_0"), val = tensor<bool, []>(false)];
311
+ tensor<bool, []> x_7_transpose_y_0 = const()[name = tensor<string, []>("x_7_transpose_y_0"), val = tensor<bool, []>(false)];
312
+ tensor<fp32, [?, 8, ?, 64]> x_7 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = softmax_0, y = squeeze_2)[name = tensor<string, []>("x_7")];
313
+ tensor<int32, [4]> var_314_shape = shape(x = x_7)[name = tensor<string, []>("op_314_shape")];
314
+ tensor<int32, []> gather_12_batch_dims_0 = const()[name = tensor<string, []>("gather_12_batch_dims_0"), val = tensor<int32, []>(0)];
315
+ tensor<bool, []> gather_12_validate_indices_0 = const()[name = tensor<string, []>("gather_12_validate_indices_0"), val = tensor<bool, []>(false)];
316
+ tensor<int32, []> select_7 = const()[name = tensor<string, []>("select_7"), val = tensor<int32, []>(0)];
317
+ tensor<int32, []> gather_12_axis_1 = const()[name = tensor<string, []>("gather_12_axis_1"), val = tensor<int32, []>(0)];
318
+ tensor<int32, []> gather_12 = gather(axis = gather_12_axis_1, batch_dims = gather_12_batch_dims_0, indices = select_7, validate_indices = gather_12_validate_indices_0, x = var_314_shape)[name = tensor<string, []>("gather_12")];
319
+ tensor<int32, []> gather_14_batch_dims_0 = const()[name = tensor<string, []>("gather_14_batch_dims_0"), val = tensor<int32, []>(0)];
320
+ tensor<bool, []> gather_14_validate_indices_0 = const()[name = tensor<string, []>("gather_14_validate_indices_0"), val = tensor<bool, []>(false)];
321
+ tensor<int32, []> select_8 = const()[name = tensor<string, []>("select_8"), val = tensor<int32, []>(2)];
322
+ tensor<int32, []> gather_14_axis_1 = const()[name = tensor<string, []>("gather_14_axis_1"), val = tensor<int32, []>(0)];
323
+ tensor<int32, []> gather_14 = gather(axis = gather_14_axis_1, batch_dims = gather_14_batch_dims_0, indices = select_8, validate_indices = gather_14_validate_indices_0, x = var_314_shape)[name = tensor<string, []>("gather_14")];
324
+ tensor<int32, []> var_331 = const()[name = tensor<string, []>("op_331"), val = tensor<int32, []>(512)];
325
+ tensor<int32, [4]> var_332 = const()[name = tensor<string, []>("op_332"), val = tensor<int32, [4]>([0, 2, 1, 3])];
326
+ tensor<int32, []> concat_6_axis_0 = const()[name = tensor<string, []>("concat_6_axis_0"), val = tensor<int32, []>(0)];
327
+ tensor<bool, []> concat_6_interleave_0 = const()[name = tensor<string, []>("concat_6_interleave_0"), val = tensor<bool, []>(false)];
328
+ tensor<int32, [3]> concat_6 = concat(axis = concat_6_axis_0, interleave = concat_6_interleave_0, values = (gather_12, gather_14, var_331))[name = tensor<string, []>("concat_6")];
329
+ tensor<fp32, [?, ?, 8, 64]> x_9 = transpose(perm = var_332, x = x_7)[name = tensor<string, []>("transpose_8")];
330
+ tensor<fp32, [?, ?, 512]> input_59 = reshape(shape = concat_6, x = x_9)[name = tensor<string, []>("input_59")];
331
+ tensor<fp32, [512]> linear_1_bias_0 = const()[name = tensor<string, []>("linear_1_bias_0"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72559808)))];
332
+ tensor<fp32, [?, ?, 512]> x_11 = linear(bias = linear_1_bias_0, weight = encoder_transformer_transformer_layers_0_self_attn_out_proj_weight, x = input_59)[name = tensor<string, []>("linear_1")];
333
+ tensor<fp32, [?, ?, 512]> var_340 = mul(x = encoder_transformer_transformer_layers_0_layer_scale_1_scale, y = x_11)[name = tensor<string, []>("op_340")];
334
+ tensor<fp32, [?, ?, 512]> input_61 = add(x = input_57, y = var_340)[name = tensor<string, []>("input_61")];
335
+ tensor<int32, [1]> input_63_axes_0 = const()[name = tensor<string, []>("input_63_axes_0"), val = tensor<int32, [1]>([-1])];
336
+ tensor<fp32, [?, ?, 512]> input_63 = layer_norm(axes = input_63_axes_0, beta = encoder_transformer_transformer_layers_0_norm2_bias, epsilon = var_178, gamma = encoder_transformer_transformer_layers_0_norm2_weight, x = input_61)[name = tensor<string, []>("input_63")];
337
+ tensor<fp32, [2048]> linear_2_bias_0 = const()[name = tensor<string, []>("linear_2_bias_0"), val = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72561920)))];
338
+ tensor<fp32, [?, ?, 2048]> var_347 = linear(bias = linear_2_bias_0, weight = encoder_transformer_transformer_layers_0_linear1_weight, x = input_63)[name = tensor<string, []>("linear_2")];
339
+ tensor<string, []> input_65_mode_0 = const()[name = tensor<string, []>("input_65_mode_0"), val = tensor<string, []>("EXACT")];
340
+ tensor<fp32, [?, ?, 2048]> input_65 = gelu(mode = input_65_mode_0, x = var_347)[name = tensor<string, []>("input_65")];
341
+ tensor<fp32, [?, ?, 512]> x_13 = linear(bias = linear_1_bias_0, weight = encoder_transformer_transformer_layers_0_linear2_weight, x = input_65)[name = tensor<string, []>("linear_3")];
342
+ tensor<fp32, [?, ?, 512]> var_353 = mul(x = encoder_transformer_transformer_layers_0_layer_scale_2_scale, y = x_13)[name = tensor<string, []>("op_353")];
343
+ tensor<fp32, [?, ?, 512]> input_67 = add(x = input_61, y = var_353)[name = tensor<string, []>("input_67")];
344
+ tensor<int32, [1]> query_axes_0 = const()[name = tensor<string, []>("query_axes_0"), val = tensor<int32, [1]>([-1])];
345
+ tensor<fp32, [?, ?, 512]> query = layer_norm(axes = query_axes_0, beta = encoder_transformer_transformer_layers_1_norm1_bias, epsilon = var_178, gamma = encoder_transformer_transformer_layers_1_norm1_weight, x = input_67)[name = tensor<string, []>("query")];
346
+ tensor<int32, [3]> var_368_shape = shape(x = query)[name = tensor<string, []>("op_368_shape")];
347
+ tensor<int32, []> gather_16_batch_dims_0 = const()[name = tensor<string, []>("gather_16_batch_dims_0"), val = tensor<int32, []>(0)];
348
+ tensor<bool, []> gather_16_validate_indices_0 = const()[name = tensor<string, []>("gather_16_validate_indices_0"), val = tensor<bool, []>(false)];
349
+ tensor<int32, []> select_9 = const()[name = tensor<string, []>("select_9"), val = tensor<int32, []>(0)];
350
+ tensor<int32, []> gather_16_axis_1 = const()[name = tensor<string, []>("gather_16_axis_1"), val = tensor<int32, []>(0)];
351
+ tensor<int32, []> gather_16 = gather(axis = gather_16_axis_1, batch_dims = gather_16_batch_dims_0, indices = select_9, validate_indices = gather_16_validate_indices_0, x = var_368_shape)[name = tensor<string, []>("gather_16")];
352
+ tensor<int32, []> gather_17_batch_dims_0 = const()[name = tensor<string, []>("gather_17_batch_dims_0"), val = tensor<int32, []>(0)];
353
+ tensor<bool, []> gather_17_validate_indices_0 = const()[name = tensor<string, []>("gather_17_validate_indices_0"), val = tensor<bool, []>(false)];
354
+ tensor<int32, []> select_10 = const()[name = tensor<string, []>("select_10"), val = tensor<int32, []>(1)];
355
+ tensor<int32, []> gather_17_axis_1 = const()[name = tensor<string, []>("gather_17_axis_1"), val = tensor<int32, []>(0)];
356
+ tensor<int32, []> gather_17 = gather(axis = gather_17_axis_1, batch_dims = gather_17_batch_dims_0, indices = select_10, validate_indices = gather_17_validate_indices_0, x = var_368_shape)[name = tensor<string, []>("gather_17")];
357
+ tensor<int32, []> concat_7_axis_0 = const()[name = tensor<string, []>("concat_7_axis_0"), val = tensor<int32, []>(0)];
358
+ tensor<bool, []> concat_7_interleave_0 = const()[name = tensor<string, []>("concat_7_interleave_0"), val = tensor<bool, []>(false)];
359
+ tensor<int32, [1]> concat_7 = concat(axis = concat_7_axis_0, interleave = concat_7_interleave_0, values = gather_16)[name = tensor<string, []>("concat_7")];
360
+ tensor<int32, []> offset_value_0 = const()[name = tensor<string, []>("offset_value_0"), val = tensor<int32, []>(0)];
361
+ tensor<int32, [?]> offset = fill(shape = concat_7, value = offset_value_0)[name = tensor<string, []>("offset")];
362
+ tensor<fp32, [?, ?, 1536]> x_15 = linear(bias = linear_0_bias_0, weight = encoder_transformer_transformer_layers_1_self_attn_in_proj_weight, x = query)[name = tensor<string, []>("linear_4")];
363
+ tensor<int32, [3]> var_374_shape = shape(x = x_15)[name = tensor<string, []>("op_374_shape")];
364
+ tensor<int32, []> gather_18_batch_dims_0 = const()[name = tensor<string, []>("gather_18_batch_dims_0"), val = tensor<int32, []>(0)];
365
+ tensor<bool, []> gather_18_validate_indices_0 = const()[name = tensor<string, []>("gather_18_validate_indices_0"), val = tensor<bool, []>(false)];
366
+ tensor<int32, []> select_11 = const()[name = tensor<string, []>("select_11"), val = tensor<int32, []>(0)];
367
+ tensor<int32, []> gather_18_axis_1 = const()[name = tensor<string, []>("gather_18_axis_1"), val = tensor<int32, []>(0)];
368
+ tensor<int32, []> gather_18 = gather(axis = gather_18_axis_1, batch_dims = gather_18_batch_dims_0, indices = select_11, validate_indices = gather_18_validate_indices_0, x = var_374_shape)[name = tensor<string, []>("gather_18")];
369
+ tensor<int32, []> gather_19_batch_dims_0 = const()[name = tensor<string, []>("gather_19_batch_dims_0"), val = tensor<int32, []>(0)];
370
+ tensor<bool, []> gather_19_validate_indices_0 = const()[name = tensor<string, []>("gather_19_validate_indices_0"), val = tensor<bool, []>(false)];
371
+ tensor<int32, []> select_12 = const()[name = tensor<string, []>("select_12"), val = tensor<int32, []>(1)];
372
+ tensor<int32, []> gather_19_axis_1 = const()[name = tensor<string, []>("gather_19_axis_1"), val = tensor<int32, []>(0)];
373
+ tensor<int32, []> gather_19 = gather(axis = gather_19_axis_1, batch_dims = gather_19_batch_dims_0, indices = select_12, validate_indices = gather_19_validate_indices_0, x = var_374_shape)[name = tensor<string, []>("gather_19")];
374
+ tensor<int32, []> var_379 = const()[name = tensor<string, []>("op_379"), val = tensor<int32, []>(64)];
375
+ tensor<int32, []> concat_8_axis_0 = const()[name = tensor<string, []>("concat_8_axis_0"), val = tensor<int32, []>(0)];
376
+ tensor<bool, []> concat_8_interleave_0 = const()[name = tensor<string, []>("concat_8_interleave_0"), val = tensor<bool, []>(false)];
377
+ tensor<int32, [5]> concat_8 = concat(axis = concat_8_axis_0, interleave = concat_8_interleave_0, values = (gather_18, gather_19, var_169, var_170, var_379))[name = tensor<string, []>("concat_8")];
378
+ tensor<fp32, [?, ?, 3, 8, 64]> x_17 = reshape(shape = concat_8, x = x_15)[name = tensor<string, []>("x_17")];
379
+ tensor<int32, [5]> var_382 = const()[name = tensor<string, []>("op_382"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])];
380
+ tensor<int32, [3]> var_384_split_sizes_0 = const()[name = tensor<string, []>("op_384_split_sizes_0"), val = tensor<int32, [3]>([1, 1, 1])];
381
+ tensor<int32, []> var_384_axis_0 = const()[name = tensor<string, []>("op_384_axis_0"), val = tensor<int32, []>(0)];
382
+ tensor<fp32, [3, ?, 8, ?, 64]> var_383 = transpose(perm = var_382, x = x_17)[name = tensor<string, []>("transpose_7")];
383
+ tensor<fp32, [1, ?, 8, ?, 64]> var_384_0, tensor<fp32, [1, ?, 8, ?, 64]> var_384_1, tensor<fp32, [1, ?, 8, ?, 64]> var_384_2 = split(axis = var_384_axis_0, split_sizes = var_384_split_sizes_0, x = var_383)[name = tensor<string, []>("op_384")];
384
+ tensor<int32, [1]> squeeze_3_axes_0 = const()[name = tensor<string, []>("squeeze_3_axes_0"), val = tensor<int32, [1]>([0])];
385
+ tensor<fp32, [?, 8, ?, 64]> squeeze_3 = squeeze(axes = squeeze_3_axes_0, x = var_384_0)[name = tensor<string, []>("squeeze_3")];
386
+ tensor<int32, [1]> squeeze_4_axes_0 = const()[name = tensor<string, []>("squeeze_4_axes_0"), val = tensor<int32, [1]>([0])];
387
+ tensor<fp32, [?, 8, ?, 64]> squeeze_4 = squeeze(axes = squeeze_4_axes_0, x = var_384_1)[name = tensor<string, []>("squeeze_4")];
388
+ tensor<int32, [1]> squeeze_5_axes_0 = const()[name = tensor<string, []>("squeeze_5_axes_0"), val = tensor<int32, [1]>([0])];
389
+ tensor<fp32, [?, 8, ?, 64]> squeeze_5 = squeeze(axes = squeeze_5_axes_0, x = var_384_2)[name = tensor<string, []>("squeeze_5")];
390
+ tensor<int32, [4]> var_388 = const()[name = tensor<string, []>("op_388"), val = tensor<int32, [4]>([0, 2, 1, 3])];
391
+ tensor<int32, [4]> var_390 = const()[name = tensor<string, []>("op_390"), val = tensor<int32, [4]>([0, 2, 1, 3])];
392
+ tensor<fp32, [?, ?, 8, 64]> q_13 = transpose(perm = var_388, x = squeeze_3)[name = tensor<string, []>("transpose_6")];
393
+ tensor<int32, [4]> var_392_shape = shape(x = q_13)[name = tensor<string, []>("op_392_shape")];
394
+ tensor<int32, []> gather_21_batch_dims_0 = const()[name = tensor<string, []>("gather_21_batch_dims_0"), val = tensor<int32, []>(0)];
395
+ tensor<bool, []> gather_21_validate_indices_0 = const()[name = tensor<string, []>("gather_21_validate_indices_0"), val = tensor<bool, []>(false)];
396
+ tensor<int32, []> select_13 = const()[name = tensor<string, []>("select_13"), val = tensor<int32, []>(0)];
397
+ tensor<int32, []> gather_21_axis_1 = const()[name = tensor<string, []>("gather_21_axis_1"), val = tensor<int32, []>(0)];
398
+ tensor<int32, []> gather_21 = gather(axis = gather_21_axis_1, batch_dims = gather_21_batch_dims_0, indices = select_13, validate_indices = gather_21_validate_indices_0, x = var_392_shape)[name = tensor<string, []>("gather_21")];
399
+ tensor<int32, []> gather_22_batch_dims_0 = const()[name = tensor<string, []>("gather_22_batch_dims_0"), val = tensor<int32, []>(0)];
400
+ tensor<bool, []> gather_22_validate_indices_0 = const()[name = tensor<string, []>("gather_22_validate_indices_0"), val = tensor<bool, []>(false)];
401
+ tensor<int32, []> select_14 = const()[name = tensor<string, []>("select_14"), val = tensor<int32, []>(1)];
402
+ tensor<int32, []> gather_22_axis_1 = const()[name = tensor<string, []>("gather_22_axis_1"), val = tensor<int32, []>(0)];
403
+ tensor<int32, []> gather_22 = gather(axis = gather_22_axis_1, batch_dims = gather_22_batch_dims_0, indices = select_14, validate_indices = gather_22_validate_indices_0, x = var_392_shape)[name = tensor<string, []>("gather_22")];
404
+ tensor<int32, []> gather_23 = const()[name = tensor<string, []>("gather_23"), val = tensor<int32, []>(8)];
405
+ tensor<int32, []> gather_24 = const()[name = tensor<string, []>("gather_24"), val = tensor<int32, []>(64)];
406
+ tensor<fp32, [32]> freqs = const()[name = tensor<string, []>("freqs"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(72570176)))];
407
+ tensor<int32, []> const_18 = const()[name = tensor<string, []>("const_18"), val = tensor<int32, []>(0)];
408
+ tensor<int32, []> const_19 = const()[name = tensor<string, []>("const_19"), val = tensor<int32, []>(1)];
409
+ tensor<int32, [?]> ts_7 = range_1d(end = gather_22, start = const_18, step = const_19)[name = tensor<string, []>("ts_7")];
410
+ tensor<int32, [?]> ts_9 = add(x = ts_7, y = offset)[name = tensor<string, []>("ts_9")];
411
+ tensor<int32, [3]> var_406 = const()[name = tensor<string, []>("op_406"), val = tensor<int32, [3]>([-1, 1, 1])];
412
+ tensor<int32, [?, 1, 1]> ts = reshape(shape = var_406, x = ts_9)[name = tensor<string, []>("ts")];
413
+ tensor<int32, []> var_409 = const()[name = tensor<string, []>("op_409"), val = tensor<int32, []>(32)];
414
+ tensor<int32, []> concat_9_axis_0 = const()[name = tensor<string, []>("concat_9_axis_0"), val = tensor<int32, []>(0)];
415
+ tensor<bool, []> concat_9_interleave_0 = const()[name = tensor<string, []>("concat_9_interleave_0"), val = tensor<bool, []>(false)];
416
+ tensor<int32, [5]> concat_9 = concat(axis = concat_9_axis_0, interleave = concat_9_interleave_0, values = (gather_21, gather_22, gather_23, var_409, var_180))[name = tensor<string, []>("concat_9")];
417
+ tensor<fp32, [?, ?, 8, 32, 2]> q_15 = reshape(shape = concat_9, x = q_13)[name = tensor<string, []>("q_15")];
418
+ tensor<fp32, [?, ?, 8, 64]> k_11 = transpose(perm = var_390, x = squeeze_4)[name = tensor<string, []>("transpose_5")];
419
+ tensor<fp32, [?, ?, 8, 32, 2]> k_13 = reshape(shape = concat_9, x = k_11)[name = tensor<string, []>("k_13")];
420
+ tensor<int32, [5]> var_416_begin_0 = const()[name = tensor<string, []>("op_416_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
421
+ tensor<int32, [5]> var_416_end_0 = const()[name = tensor<string, []>("op_416_end_0"), val = tensor<int32, [5]>([0, 0, 8, 32, 1])];
422
+ tensor<bool, [5]> var_416_end_mask_0 = const()[name = tensor<string, []>("op_416_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
423
+ tensor<bool, [5]> var_416_squeeze_mask_0 = const()[name = tensor<string, []>("op_416_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
424
+ tensor<fp32, [?, ?, 8, 32]> var_416 = slice_by_index(begin = var_416_begin_0, end = var_416_end_0, end_mask = var_416_end_mask_0, squeeze_mask = var_416_squeeze_mask_0, x = q_15)[name = tensor<string, []>("op_416")];
425
+ tensor<int32, [5]> var_418_begin_0 = const()[name = tensor<string, []>("op_418_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
426
+ tensor<int32, [5]> var_418_end_0 = const()[name = tensor<string, []>("op_418_end_0"), val = tensor<int32, [5]>([0, 0, 8, 32, 2])];
427
+ tensor<bool, [5]> var_418_end_mask_0 = const()[name = tensor<string, []>("op_418_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
428
+ tensor<bool, [5]> var_418_squeeze_mask_0 = const()[name = tensor<string, []>("op_418_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
429
+ tensor<fp32, [?, ?, 8, 32]> var_418 = slice_by_index(begin = var_418_begin_0, end = var_418_end_0, end_mask = var_418_end_mask_0, squeeze_mask = var_418_squeeze_mask_0, x = q_15)[name = tensor<string, []>("op_418")];
430
+ tensor<int32, [5]> var_420_begin_0 = const()[name = tensor<string, []>("op_420_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
431
+ tensor<int32, [5]> var_420_end_0 = const()[name = tensor<string, []>("op_420_end_0"), val = tensor<int32, [5]>([0, 0, 8, 32, 1])];
432
+ tensor<bool, [5]> var_420_end_mask_0 = const()[name = tensor<string, []>("op_420_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
433
+ tensor<bool, [5]> var_420_squeeze_mask_0 = const()[name = tensor<string, []>("op_420_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
434
+ tensor<fp32, [?, ?, 8, 32]> var_420 = slice_by_index(begin = var_420_begin_0, end = var_420_end_0, end_mask = var_420_end_mask_0, squeeze_mask = var_420_squeeze_mask_0, x = k_13)[name = tensor<string, []>("op_420")];
435
+ tensor<int32, [5]> var_422_begin_0 = const()[name = tensor<string, []>("op_422_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
436
+ tensor<int32, [5]> var_422_end_0 = const()[name = tensor<string, []>("op_422_end_0"), val = tensor<int32, [5]>([0, 0, 8, 32, 2])];
437
+ tensor<bool, [5]> var_422_end_mask_0 = const()[name = tensor<string, []>("op_422_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
438
+ tensor<bool, [5]> var_422_squeeze_mask_0 = const()[name = tensor<string, []>("op_422_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
439
+ tensor<fp32, [?, ?, 8, 32]> var_422 = slice_by_index(begin = var_422_begin_0, end = var_422_end_0, end_mask = var_422_end_mask_0, squeeze_mask = var_422_squeeze_mask_0, x = k_13)[name = tensor<string, []>("op_422")];
440
+ tensor<string, []> ts_promoted_dtype_0 = const()[name = tensor<string, []>("ts_promoted_dtype_0"), val = tensor<string, []>("fp32")];
441
+ tensor<fp32, [?, 1, 1]> ts_promoted = cast(dtype = ts_promoted_dtype_0, x = ts)[name = tensor<string, []>("cast_32")];
442
+ tensor<fp32, [?, 1, 32]> var_424 = mul(x = freqs, y = ts_promoted)[name = tensor<string, []>("op_424")];
443
+ tensor<fp32, [?, 1, 32]> rotr = cos(x = var_424)[name = tensor<string, []>("rotr")];
444
+ tensor<fp32, [?, 1, 32]> roti = sin(x = var_424)[name = tensor<string, []>("roti")];
445
+ tensor<fp32, [?, ?, 8, 32]> var_428 = mul(x = var_416, y = rotr)[name = tensor<string, []>("op_428")];
446
+ tensor<fp32, [?, ?, 8, 32]> var_429 = mul(x = var_418, y = roti)[name = tensor<string, []>("op_429")];
447
+ tensor<fp32, [?, ?, 8, 32]> qor_5 = sub(x = var_428, y = var_429)[name = tensor<string, []>("qor_5")];
448
+ tensor<fp32, [?, ?, 8, 32]> var_431 = mul(x = var_416, y = roti)[name = tensor<string, []>("op_431")];
449
+ tensor<fp32, [?, ?, 8, 32]> var_432 = mul(x = var_418, y = rotr)[name = tensor<string, []>("op_432")];
450
+ tensor<fp32, [?, ?, 8, 32]> qoi_5 = add(x = var_431, y = var_432)[name = tensor<string, []>("qoi_5")];
451
+ tensor<fp32, [?, ?, 8, 32]> var_434 = mul(x = var_420, y = rotr)[name = tensor<string, []>("op_434")];
452
+ tensor<fp32, [?, ?, 8, 32]> var_435 = mul(x = var_422, y = roti)[name = tensor<string, []>("op_435")];
453
+ tensor<fp32, [?, ?, 8, 32]> kor_5 = sub(x = var_434, y = var_435)[name = tensor<string, []>("kor_5")];
454
+ tensor<fp32, [?, ?, 8, 32]> var_437 = mul(x = var_420, y = roti)[name = tensor<string, []>("op_437")];
455
+ tensor<fp32, [?, ?, 8, 32]> var_438 = mul(x = var_422, y = rotr)[name = tensor<string, []>("op_438")];
456
+ tensor<fp32, [?, ?, 8, 32]> koi_5 = add(x = var_437, y = var_438)[name = tensor<string, []>("koi_5")];
457
+ tensor<int32, []> qo_axis_0 = const()[name = tensor<string, []>("qo_axis_0"), val = tensor<int32, []>(-1)];
458
+ tensor<fp32, [?, ?, 8, 32, 2]> qo = stack(axis = qo_axis_0, values = (qor_5, qoi_5))[name = tensor<string, []>("qo")];
459
+ tensor<int32, []> ko_axis_0 = const()[name = tensor<string, []>("ko_axis_0"), val = tensor<int32, []>(-1)];
460
+ tensor<fp32, [?, ?, 8, 32, 2]> ko = stack(axis = ko_axis_0, values = (kor_5, koi_5))[name = tensor<string, []>("ko")];
461
+ tensor<int32, []> concat_11_axis_0 = const()[name = tensor<string, []>("concat_11_axis_0"), val = tensor<int32, []>(0)];
462
+ tensor<bool, []> concat_11_interleave_0 = const()[name = tensor<string, []>("concat_11_interleave_0"), val = tensor<bool, []>(false)];
463
+ tensor<int32, [4]> concat_11 = concat(axis = concat_11_axis_0, interleave = concat_11_interleave_0, values = (gather_21, gather_22, gather_23, gather_24))[name = tensor<string, []>("concat_11")];
464
+ tensor<fp32, [?, ?, 8, 64]> q_17 = reshape(shape = concat_11, x = qo)[name = tensor<string, []>("q_17")];
465
+ tensor<fp32, [?, ?, 8, 64]> k = reshape(shape = concat_11, x = ko)[name = tensor<string, []>("k")];
466
+ tensor<int32, [4]> var_455 = const()[name = tensor<string, []>("op_455"), val = tensor<int32, [4]>([0, 2, 1, 3])];
467
+ tensor<int32, [4]> var_457 = const()[name = tensor<string, []>("op_457"), val = tensor<int32, [4]>([0, 2, 1, 3])];
468
+ tensor<fp32, [?, 8, ?, 64]> keys = transpose(perm = var_457, x = k)[name = tensor<string, []>("transpose_3")];
469
+ tensor<int32, [4]> var_459_shape = shape(x = keys)[name = tensor<string, []>("op_459_shape")];
470
+ tensor<int32, []> gather_26_batch_dims_0 = const()[name = tensor<string, []>("gather_26_batch_dims_0"), val = tensor<int32, []>(0)];
471
+ tensor<bool, []> gather_26_validate_indices_0 = const()[name = tensor<string, []>("gather_26_validate_indices_0"), val = tensor<bool, []>(false)];
472
+ tensor<int32, []> select_15 = const()[name = tensor<string, []>("select_15"), val = tensor<int32, []>(0)];
473
+ tensor<int32, []> gather_26_axis_1 = const()[name = tensor<string, []>("gather_26_axis_1"), val = tensor<int32, []>(0)];
474
+ tensor<int32, []> gather_26 = gather(axis = gather_26_axis_1, batch_dims = gather_26_batch_dims_0, indices = select_15, validate_indices = gather_26_validate_indices_0, x = var_459_shape)[name = tensor<string, []>("gather_26")];
475
+ tensor<int32, []> gather_27_batch_dims_0 = const()[name = tensor<string, []>("gather_27_batch_dims_0"), val = tensor<int32, []>(0)];
476
+ tensor<bool, []> gather_27_validate_indices_0 = const()[name = tensor<string, []>("gather_27_validate_indices_0"), val = tensor<bool, []>(false)];
477
+ tensor<int32, []> select_16 = const()[name = tensor<string, []>("select_16"), val = tensor<int32, []>(2)];
478
+ tensor<int32, []> gather_27_axis_1 = const()[name = tensor<string, []>("gather_27_axis_1"), val = tensor<int32, []>(0)];
479
+ tensor<int32, []> gather_27 = gather(axis = gather_27_axis_1, batch_dims = gather_27_batch_dims_0, indices = select_16, validate_indices = gather_27_validate_indices_0, x = var_459_shape)[name = tensor<string, []>("gather_27")];
480
+ tensor<int32, []> const_20 = const()[name = tensor<string, []>("const_20"), val = tensor<int32, []>(0)];
481
+ tensor<int32, []> const_21 = const()[name = tensor<string, []>("const_21"), val = tensor<int32, []>(1)];
482
+ tensor<int32, [?]> positions = range_1d(end = gather_27, start = const_20, step = const_21)[name = tensor<string, []>("positions")];
483
+ tensor<int32, []> concat_13_axis_0 = const()[name = tensor<string, []>("concat_13_axis_0"), val = tensor<int32, []>(0)];
484
+ tensor<bool, []> concat_13_interleave_0 = const()[name = tensor<string, []>("concat_13_interleave_0"), val = tensor<bool, []>(false)];
485
+ tensor<int32, [2]> concat_13 = concat(axis = concat_13_axis_0, interleave = concat_13_interleave_0, values = (gather_26, var_171))[name = tensor<string, []>("concat_13")];
486
+ tensor<int32, [1]> expand_dims_1_axes_0 = const()[name = tensor<string, []>("expand_dims_1_axes_0"), val = tensor<int32, [1]>([0])];
487
+ tensor<int32, [1, ?]> expand_dims_1 = expand_dims(axes = expand_dims_1_axes_0, x = positions)[name = tensor<string, []>("expand_dims_1")];
488
+ tensor<int32, [2]> shape_1 = shape(x = expand_dims_1)[name = tensor<string, []>("shape_1")];
489
+ tensor<int32, []> equal_1_y_0 = const()[name = tensor<string, []>("equal_1_y_0"), val = tensor<int32, []>(-1)];
490
+ tensor<bool, [2]> equal_1 = equal(x = concat_13, y = equal_1_y_0)[name = tensor<string, []>("equal_1")];
491
+ tensor<int32, [2]> select_1 = select(a = shape_1, b = concat_13, cond = equal_1)[name = tensor<string, []>("select_1")];
492
+ tensor<int32, [2]> real_div_1 = real_div(x = select_1, y = shape_1)[name = tensor<string, []>("real_div_1")];
493
+ tensor<int32, [?, ?]> pos_k_5 = tile(reps = real_div_1, x = expand_dims_1)[name = tensor<string, []>("pos_k_5")];
494
+ tensor<int32, [1]> pos_k_axes_0 = const()[name = tensor<string, []>("pos_k_axes_0"), val = tensor<int32, [1]>([1])];
495
+ tensor<int32, [?, 1, ?]> pos_k = expand_dims(axes = pos_k_axes_0, x = pos_k_5)[name = tensor<string, []>("pos_k")];
496
+ tensor<int32, [3]> var_466 = const()[name = tensor<string, []>("op_466"), val = tensor<int32, [3]>([-1, 1, 1])];
497
+ tensor<int32, [?, 1, 1]> var_467 = reshape(shape = var_466, x = offset)[name = tensor<string, []>("op_467")];
498
+ tensor<int32, []> const_22 = const()[name = tensor<string, []>("const_22"), val = tensor<int32, []>(0)];
499
+ tensor<int32, []> const_23 = const()[name = tensor<string, []>("const_23"), val = tensor<int32, []>(1)];
500
+ tensor<int32, [?]> var_468 = range_1d(end = gather_17, start = const_22, step = const_23)[name = tensor<string, []>("op_468")];
501
+ tensor<int32, [2]> var_469 = const()[name = tensor<string, []>("op_469"), val = tensor<int32, [2]>([-1, 1])];
502
+ tensor<int32, [?, 1]> var_470 = reshape(shape = var_469, x = var_468)[name = tensor<string, []>("op_470")];
503
+ tensor<int32, [?, ?, 1]> pos_q = add(x = var_467, y = var_470)[name = tensor<string, []>("pos_q")];
504
+ tensor<int32, [?, ?, ?]> delta = sub(x = pos_q, y = pos_k)[name = tensor<string, []>("delta")];
505
+ tensor<bool, [?, 1, ?]> var_473 = greater_equal(x = pos_k, y = var_163)[name = tensor<string, []>("op_473")];
506
+ tensor<bool, [?, ?, ?]> var_474 = greater_equal(x = delta, y = var_163)[name = tensor<string, []>("op_474")];
507
+ tensor<bool, [?, ?, ?]> attn_bias_7 = logical_and(x = var_473, y = var_474)[name = tensor<string, []>("attn_bias_7")];
508
+ tensor<bool, [?, ?, ?]> var_476 = less(x = delta, y = var_175)[name = tensor<string, []>("op_476")];
509
+ tensor<bool, [?, ?, ?]> attn_bias_9 = logical_and(x = attn_bias_7, y = var_476)[name = tensor<string, []>("attn_bias_9")];
510
+ tensor<int32, [1]> attn_bias_axes_0 = const()[name = tensor<string, []>("attn_bias_axes_0"), val = tensor<int32, [1]>([1])];
511
+ tensor<bool, [?, 1, ?, ?]> attn_bias = expand_dims(axes = attn_bias_axes_0, x = attn_bias_9)[name = tensor<string, []>("attn_bias")];
512
+ tensor<string, []> cast_28_dtype_0 = const()[name = tensor<string, []>("cast_28_dtype_0"), val = tensor<string, []>("fp32")];
513
+ tensor<fp32, []> sub_1_x_0 = const()[name = tensor<string, []>("sub_1_x_0"), val = tensor<fp32, []>(0x1p+0)];
514
+ tensor<fp32, [?, 1, ?, ?]> cast_28 = cast(dtype = cast_28_dtype_0, x = attn_bias)[name = tensor<string, []>("cast_31")];
515
+ tensor<fp32, [?, 1, ?, ?]> sub_1 = sub(x = sub_1_x_0, y = cast_28)[name = tensor<string, []>("sub_1")];
516
+ tensor<fp32, []> mul_2_x_0 = const()[name = tensor<string, []>("mul_2_x_0"), val = tensor<fp32, []>(-0x1.d4cp+14)];
517
+ tensor<fp32, [?, 1, ?, ?]> mul_2 = mul(x = mul_2_x_0, y = sub_1)[name = tensor<string, []>("mul_2")];
518
+ tensor<fp32, []> mul_3_y_0 = const()[name = tensor<string, []>("mul_3_y_0"), val = tensor<fp32, []>(0x1p-3)];
519
+ tensor<fp32, [?, 8, ?, 64]> q = transpose(perm = var_455, x = q_17)[name = tensor<string, []>("transpose_4")];
520
+ tensor<fp32, [?, 8, ?, 64]> mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor<string, []>("mul_3")];
521
+ tensor<bool, []> matmul_1_transpose_y_0 = const()[name = tensor<string, []>("matmul_1_transpose_y_0"), val = tensor<bool, []>(true)];
522
+ tensor<bool, []> matmul_1_transpose_x_0 = const()[name = tensor<string, []>("matmul_1_transpose_x_0"), val = tensor<bool, []>(false)];
523
+ tensor<fp32, [?, 8, ?, ?]> matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = keys)[name = tensor<string, []>("matmul_1")];
524
+ tensor<fp32, [?, 8, ?, ?]> add_1 = add(x = matmul_1, y = mul_2)[name = tensor<string, []>("add_1")];
525
+ tensor<int32, []> softmax_1_axis_0 = const()[name = tensor<string, []>("softmax_1_axis_0"), val = tensor<int32, []>(-1)];
526
+ tensor<fp32, [?, 8, ?, ?]> softmax_1 = softmax(axis = softmax_1_axis_0, x = add_1)[name = tensor<string, []>("softmax_1")];
527
+ tensor<bool, []> x_19_transpose_x_0 = const()[name = tensor<string, []>("x_19_transpose_x_0"), val = tensor<bool, []>(false)];
528
+ tensor<bool, []> x_19_transpose_y_0 = const()[name = tensor<string, []>("x_19_transpose_y_0"), val = tensor<bool, []>(false)];
529
+ tensor<fp32, [?, 8, ?, 64]> x_19 = matmul(transpose_x = x_19_transpose_x_0, transpose_y = x_19_transpose_y_0, x = softmax_1, y = squeeze_5)[name = tensor<string, []>("x_19")];
530
+ tensor<int32, [4]> var_481_shape = shape(x = x_19)[name = tensor<string, []>("op_481_shape")];
531
+ tensor<int32, []> gather_28_batch_dims_0 = const()[name = tensor<string, []>("gather_28_batch_dims_0"), val = tensor<int32, []>(0)];
532
+ tensor<bool, []> gather_28_validate_indices_0 = const()[name = tensor<string, []>("gather_28_validate_indices_0"), val = tensor<bool, []>(false)];
533
+ tensor<int32, []> select_17 = const()[name = tensor<string, []>("select_17"), val = tensor<int32, []>(0)];
534
+ tensor<int32, []> gather_28_axis_1 = const()[name = tensor<string, []>("gather_28_axis_1"), val = tensor<int32, []>(0)];
535
+ tensor<int32, []> gather_28 = gather(axis = gather_28_axis_1, batch_dims = gather_28_batch_dims_0, indices = select_17, validate_indices = gather_28_validate_indices_0, x = var_481_shape)[name = tensor<string, []>("gather_28")];
536
+ tensor<int32, []> gather_30_batch_dims_0 = const()[name = tensor<string, []>("gather_30_batch_dims_0"), val = tensor<int32, []>(0)];
537
+ tensor<bool, []> gather_30_validate_indices_0 = const()[name = tensor<string, []>("gather_30_validate_indices_0"), val = tensor<bool, []>(false)];
538
+ tensor<int32, []> select_18 = const()[name = tensor<string, []>("select_18"), val = tensor<int32, []>(2)];
539
+ tensor<int32, []> gather_30_axis_1 = const()[name = tensor<string, []>("gather_30_axis_1"), val = tensor<int32, []>(0)];
540
+ tensor<int32, []> gather_30 = gather(axis = gather_30_axis_1, batch_dims = gather_30_batch_dims_0, indices = select_18, validate_indices = gather_30_validate_indices_0, x = var_481_shape)[name = tensor<string, []>("gather_30")];
541
+ tensor<int32, []> var_498 = const()[name = tensor<string, []>("op_498"), val = tensor<int32, []>(512)];
542
+ tensor<int32, [4]> var_499 = const()[name = tensor<string, []>("op_499"), val = tensor<int32, [4]>([0, 2, 1, 3])];
543
+ tensor<int32, []> concat_14_axis_0 = const()[name = tensor<string, []>("concat_14_axis_0"), val = tensor<int32, []>(0)];
544
+ tensor<bool, []> concat_14_interleave_0 = const()[name = tensor<string, []>("concat_14_interleave_0"), val = tensor<bool, []>(false)];
545
+ tensor<int32, [3]> concat_14 = concat(axis = concat_14_axis_0, interleave = concat_14_interleave_0, values = (gather_28, gather_30, var_498))[name = tensor<string, []>("concat_14")];
546
+ tensor<fp32, [?, ?, 8, 64]> x_21 = transpose(perm = var_499, x = x_19)[name = tensor<string, []>("transpose_2")];
547
+ tensor<fp32, [?, ?, 512]> input_69 = reshape(shape = concat_14, x = x_21)[name = tensor<string, []>("input_69")];
548
+ tensor<fp32, [?, ?, 512]> x_23 = linear(bias = linear_1_bias_0, weight = encoder_transformer_transformer_layers_1_self_attn_out_proj_weight, x = input_69)[name = tensor<string, []>("linear_5")];
549
+ tensor<fp32, [?, ?, 512]> var_507 = mul(x = encoder_transformer_transformer_layers_1_layer_scale_1_scale, y = x_23)[name = tensor<string, []>("op_507")];
550
+ tensor<fp32, [?, ?, 512]> input_71 = add(x = input_67, y = var_507)[name = tensor<string, []>("input_71")];
551
+ tensor<int32, [1]> input_73_axes_0 = const()[name = tensor<string, []>("input_73_axes_0"), val = tensor<int32, [1]>([-1])];
552
+ tensor<fp32, [?, ?, 512]> input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_transformer_transformer_layers_1_norm2_bias, epsilon = var_178, gamma = encoder_transformer_transformer_layers_1_norm2_weight, x = input_71)[name = tensor<string, []>("input_73")];
553
+ tensor<fp32, [?, ?, 2048]> var_514 = linear(bias = linear_2_bias_0, weight = encoder_transformer_transformer_layers_1_linear1_weight, x = input_73)[name = tensor<string, []>("linear_6")];
554
+ tensor<string, []> input_75_mode_0 = const()[name = tensor<string, []>("input_75_mode_0"), val = tensor<string, []>("EXACT")];
555
+ tensor<fp32, [?, ?, 2048]> input_75 = gelu(mode = input_75_mode_0, x = var_514)[name = tensor<string, []>("input_75")];
556
+ tensor<fp32, [?, ?, 512]> x_25 = linear(bias = linear_1_bias_0, weight = encoder_transformer_transformer_layers_1_linear2_weight, x = input_75)[name = tensor<string, []>("linear_7")];
557
+ tensor<fp32, [?, ?, 512]> var_520 = mul(x = encoder_transformer_transformer_layers_1_layer_scale_2_scale, y = x_25)[name = tensor<string, []>("op_520")];
558
+ tensor<fp32, [?, ?, 512]> z = add(x = input_71, y = var_520)[name = tensor<string, []>("z")];
559
+ tensor<int32, [3]> x_perm_0 = const()[name = tensor<string, []>("x_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
560
+ tensor<int32, []> var_526 = const()[name = tensor<string, []>("op_526"), val = tensor<int32, []>(-1)];
561
+ tensor<int32, [3]> var_533_begin_0 = const()[name = tensor<string, []>("op_533_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
562
+ tensor<int32, [3]> var_533_end_0 = const()[name = tensor<string, []>("op_533_end_0"), val = tensor<int32, [3]>([0, 512, 1])];
563
+ tensor<bool, [3]> var_533_end_mask_0 = const()[name = tensor<string, []>("op_533_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
564
+ tensor<fp32, [?, 512, ?]> x = transpose(perm = x_perm_0, x = z)[name = tensor<string, []>("transpose_1")];
565
+ tensor<fp32, [?, 512, 1]> var_533 = slice_by_index(begin = var_533_begin_0, end = var_533_end_0, end_mask = var_533_end_mask_0, x = x)[name = tensor<string, []>("op_533")];
566
+ tensor<int32, [3]> concat_15 = const()[name = tensor<string, []>("concat_15"), val = tensor<int32, [3]>([-1, -1, 16])];
567
+ tensor<int32, [3]> shape_2 = shape(x = var_533)[name = tensor<string, []>("shape_2")];
568
+ tensor<bool, [3]> equal_2 = const()[name = tensor<string, []>("equal_2"), val = tensor<bool, [3]>([true, true, false])];
569
+ tensor<int32, [3]> select_2 = select(a = shape_2, b = concat_15, cond = equal_2)[name = tensor<string, []>("select_2")];
570
+ tensor<int32, [3]> real_div_2 = real_div(x = select_2, y = shape_2)[name = tensor<string, []>("real_div_2")];
571
+ tensor<fp32, [?, ?, ?]> init = tile(reps = real_div_2, x = var_533)[name = tensor<string, []>("init")];
572
+ tensor<bool, []> input_interleave_0 = const()[name = tensor<string, []>("input_interleave_0"), val = tensor<bool, []>(false)];
573
+ tensor<fp32, [?, ?, ?]> input = concat(axis = var_526, interleave = input_interleave_0, values = (init, x))[name = tensor<string, []>("input")];
574
+ tensor<string, []> emb_pad_type_0 = const()[name = tensor<string, []>("emb_pad_type_0"), val = tensor<string, []>("valid")];
575
+ tensor<int32, [1]> emb_strides_0 = const()[name = tensor<string, []>("emb_strides_0"), val = tensor<int32, [1]>([16])];
576
+ tensor<int32, [2]> emb_pad_0 = const()[name = tensor<string, []>("emb_pad_0"), val = tensor<int32, [2]>([0, 0])];
577
+ tensor<int32, [1]> emb_dilations_0 = const()[name = tensor<string, []>("emb_dilations_0"), val = tensor<int32, [1]>([1])];
578
+ tensor<int32, []> emb_groups_0 = const()[name = tensor<string, []>("emb_groups_0"), val = tensor<int32, []>(1)];
579
+ tensor<fp32, [?, 512, ?]> emb = conv(dilations = emb_dilations_0, groups = emb_groups_0, pad = emb_pad_0, pad_type = emb_pad_type_0, strides = emb_strides_0, weight = downsample_conv_conv_weight, x = input)[name = tensor<string, []>("emb")];
580
+ tensor<int32, [3]> var_546_perm_0 = const()[name = tensor<string, []>("op_546_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
581
+ 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, []>(72570368)))];
582
+ tensor<fp32, [?, ?, 512]> var_546 = transpose(perm = var_546_perm_0, x = emb)[name = tensor<string, []>("transpose_0")];
583
+ tensor<fp32, [?, ?, 1024]> conditioning = linear(bias = linear_8_bias_0, weight = speaker_proj_weight, x = var_546)[name = tensor<string, []>("linear_8")];
584
+ } -> (conditioning);
585
+ }
mimi_encoder.mlmodelc/weights/weight.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f1b85120d97529b62e947b72658be7a3002070828bd5b6b9fed78ea484b6290c
3
+ size 72574528
mimi_encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2a5c81586b6c033c5cba0e4c393fb1966bbf524f2404ec9a8e7b47261940d1df
3
+ size 77354
mimi_encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f1b85120d97529b62e947b72658be7a3002070828bd5b6b9fed78ea484b6290c
3
+ size 72574528
mimi_encoder.mlpackage/Manifest.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "fileFormatVersion": "1.0.0",
3
+ "itemInfoEntries": {
4
+ "0D853A71-3339-4D97-893A-CBCBABA96229": {
5
+ "author": "com.apple.CoreML",
6
+ "description": "CoreML Model Specification",
7
+ "name": "model.mlmodel",
8
+ "path": "com.apple.CoreML/model.mlmodel"
9
+ },
10
+ "E1672A33-053D-4924-9DA9-220138DB3728": {
11
+ "author": "com.apple.CoreML",
12
+ "description": "CoreML Model Weights",
13
+ "name": "weights",
14
+ "path": "com.apple.CoreML/weights"
15
+ }
16
+ },
17
+ "rootModelIdentifier": "0D853A71-3339-4D97-893A-CBCBABA96229"
18
+ }